Daily Digest — 2026-05-30
218 items · 10 research labs, 200 arxiv papers, 8 industry media
🏛️ Research Labs (10)
Boston Children’s uses AI to unlock new diagnoses
Boston Children’s Hospital implemented an enterprise AI layer, including a secure internal ChatGPT environment, to integrate AI across clinical, research, and administrative workflows. The system enabled rapid deployment of AI tools, automating repetitive tasks, synthesizing medical literature, and supporting rare disease diagnosis. Results include over 40 previously unresolved rare disease diagnoses, 60,000 hours saved across workflows, $7M+ in redeployed labor, and 50+ operational automations. The hospital’s AI strategy focuses on deeper clinical integration, broader adoption, and collaboration with OpenAI, positioning AI as a core component of medical practice.
enterprise ai layerrare disease diagnosisoperational automationsclinical decision supportchatgpt environment
How Braintrust turns customer requests into code with Codex
Braintrust leverages OpenAI's Codex with GPT‑5.5 to streamline customer feature request implementation, reducing development cycles from backlog prioritization to real-time preview branches. By integrating Codex into their workflow, engineers copy-paste requests, generate preview branches, and iterate with customers in minutes, enhancing feedback loops and experimentation speed. Within one month, 50% of the team adopted Codex, citing its ability to maintain terminal output speed and reduce setup overhead for new ideas. This approach shifts from step-by-step prompting to defining problems in sandbox environments, accelerating ideation and problem-solving.
codexgpt‑5.5preview branchessandbox environmentterminal output
Strengthening societal resilience with Rosalind Biodefense
OpenAI introduces Rosalind Biodefense, a program providing trusted developers with access to GPT-Rosalind, a frontier reasoning model for life sciences, to build biodefense and pandemic preparedness tools. The initiative includes expanded access for select U.S. government and allied partners, focusing on applications like epidemiological modeling, early detection, and medical countermeasure development. Initial collaborations include Fourth Eon Biosecurity, Lawrence Livermore National Laboratory, and Johns Hopkins Applied Physics Laboratory, demonstrating AI's role in enhancing screening systems, biopreparedness, and protein-engineering platforms.
gpt-rosalindbiodefensepandemic preparednessepidemiological modelingprotein-engineering
A shared playbook for trustworthy third party evaluations
OpenAI proposes a framework for trustworthy third-party evaluations of frontier AI models, emphasizing the critical role of evaluation harnesses in accurately assessing capabilities and safeguards. The method distinguishes between capability elicitation, safeguard performance, and comparative evaluations, while highlighting risks like reward hacking, refusals, and contamination. Results demonstrate that harness design (e.g., state preservation, tool use) and compute budgets significantly impact measured performance, as shown in GPT-5.5 cyber evaluations where token budget increases improved performance by up to 59%. The framework recommends explicit reporting of claim types, harness choices, and validity checks to ensure interpretable evaluations.
evaluation harnesscapability elicitationreward hackingsafeguard robustnesscontamination detection
How Endava builds an agentic organization with Codex
Endava demonstrates how OpenAI's Codex enables agentic organizations by codifying senior engineering expertise into scalable workflows. The firm employs Codex as a general desktop agent across software delivery stages—requirements analysis, design, specifications, development, and operations—reducing multi-week processes to hours. Results include 10x faster requirements specification (2 hours vs. 1-2 weeks) and live architectural diagram generation during client sessions. Key mechanisms include in-context knowledge transfer from senior to junior engineers and parallelized mentorship via AI-mediated best practices.
agentic organizationrequirements analysisknowledge transferarchitectural diagramdesktop agent
MUFG aims to become AI-native with OpenAI
Mitsubishi UFJ Financial Group (MUFG) deployed ChatGPT Enterprise to 35,000 employees, aiming to transform financial operations through generative AI. The implementation involved enterprise-grade security, mandatory e-learning (100% participation), and custom GPT workshops, resulting in 1,800 department-specific GPTs created within four months. Early results show 20-30% workload reduction in research tasks and the development of 'AI bankers' for specialized workflows. MUFG also plans AI-driven customer interfaces, including a conversational AI concierge for its digital bank, emutt.
chatgpt enterprisegenerative aicustom gptsai-nativerobo-advisory
Profiling in PyTorch (Part 1): A Beginner's Guide to torch.profiler
The article introduces torch.profiler as a tool for optimizing PyTorch operations through systematic profiling, focusing on matrix multiplication and bias addition as a foundational case study. It demonstrates profiling setup, interpretation of CPU/GPU trace disparities, and the impact of warmup runs on measurement accuracy. Results show that small matrix operations are overhead-bound (23.104μs GPU vs 2.314ms CPU), while larger matrices (4096x4096) shift to compute-bound regimes (4.495ms GPU vs 4.908ms CPU). The trace analysis reveals kernel launch delays and buffer allocation patterns using Perfetto UI.
torch.profileroverhead-boundcompute-boundperfetto uicuda kernels
Take our I/O 2026 quiz, vibe coded in Google AI Studio.
The Google AI Blog introduces a quiz built using Google AI Studio to engage developers with I/O 2026 announcements. The quiz demonstrates the platform's accessibility, as it was created by a non-developer using Gemini-generated prompts and uploaded sources. Google AI Studio, powered by the Antigravity coding agent, enables users to leverage Gemini models for creative projects without extensive coding expertise. The article highlights the tool's potential for prompt refinement and project realization, inviting readers to explore its capabilities.
geminipromptantigravitystudioquiz
9 demos of Gemini Omni and Gemini 3.5 in action
Google introduced Gemini Omni and Gemini 3.5 Flash, showcasing multimodal capabilities and agentic task execution. Gemini Omni enables video editing through natural language prompts, maintaining consistency and physics across edits, while generating high-quality outputs from combined inputs like images, audio, and text. Gemini 3.5 Flash excels in long-horizon agentic tasks, leveraging Antigravity for multi-step workflows and coding, and powers interactive UIs and intelligent experiences in Search and Gemini apps. Both models are integrated into Google’s ecosystem, with Gemini Omni available via YouTube Shorts and Gemini 3.5 Flash accessible through APIs and AI Studio.
multimodalagentic tasksnatural languageantigravitygemini api
Check out real-life AI prototypes from the Futures Lab.
The Google-funded Futures Lab at University of Waterloo developed three AI-powered educational prototypes through an eight-week interdisciplinary workshop. Kanji Garden employs generative AI for Japanese language acquisition via contextual stories, SignFluent provides real-time computer vision feedback for ASL learners, and MuscleMemory uses pose estimation for calisthenics form correction. Led by Dr. Edith Law, the program emphasized user-centered design, with teams reporting insights on accessibility integration (SignFluent), applied communication skills (MuscleMemory), and pedagogical UX design (Kanji Garden).
generative aicomputer visionpose estimationuser-centered designpedagogical ux
📜 arXiv Papers (200)
Physics Is All You Need? A Case Study in Physicist-Supervised AI Development of Scientific Software
The study examines AI-assisted scientific software development through a 12-day case study where a physicist supervised Claude Code models (Sonnet and Opus) to build CLAX-PT, a JAX-based differentiable one-loop perturbation theory module. Over 57 sessions, the agent autonomously resolved 10/15 issues via test iteration, while 2 required domain expertise. Three failures involved misdiagnosing symptoms as root causes, persisting for 33 sessions until physics-informed intervention. Key supervision practices included diverse parameter testing, shared changelogs, and prohibiting unphysical patches. Results indicate current agents lack capacity for architectural redesign or explanatory correctness, with supervision design critically determining output trustworthiness.
differentiable physicsoracle testingsupervision eventsparameter calibrationexplanatory correctness
VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion
VideoMLA introduces Multi-Head Latent Attention (MLA) for minute-scale autoregressive video diffusion, replacing per-head KV caches with a shared low-rank content latent and decoupled 3D-RoPE positional key, reducing per-token KV memory by 92.7%. Contrary to spectral assumptions in language models, pretrained video attention is not low-rank, yet MLA retains quality despite high reconstruction error predictions. The bottleneck determines effective rank, with both spectral and random initialization occupying the full rank budget. On VBench, VideoMLA matches short-horizon baselines, excels at long horizons, and improves throughput by 1.23x on a B200 GPU.
multi-head latent attentionkv cache3d-ropevideo diffusionlow-rank approximation
LLMSurgeon: Diagnosing Data Mixture of Large Language Models
LLMSurgeon introduces Data Mixture Surgery (DMS), a framework for estimating the domain-level distribution of pretraining data in Large Language Models (LLMs) solely from generated text. The method casts DMS as an inverse problem under label-shift assumptions, employing a calibrated soft confusion matrix to correct systematic domain confusion and recover the latent mixture prior. Evaluated on LLMScan, a verifiable suite of open-source LLMs with transparent pretraining mixtures, LLMSurgeon achieves high fidelity in recovering domain distributions. This work enables post-hoc auditing of LLM pretraining data composition without access to training data.
data mixture surgerylarge language modelsinverse problemlabel-shiftsoft confusion matrix
SchGen: PCB Schematic Generation with Semantic-Grounded Code Representations
SchGen introduces the first LLM for generating PCB schematics from natural-language prompts by developing a semantically grounded code representation that simplifies schematic generation into a semantics-driven matching task. The method addresses challenges of verbose, tool-specific formats by encoding schematic primitives with relative placement and pin-name-based wiring, supported by a large-scale dataset created via human-agent collaboration. Results demonstrate SchGen's superiority over alternative representations and larger general-purpose LLMs in wire connectivity accuracy and functional correctness, emphasizing representation design's critical role in hardware generation tasks.
pcb schematicsemantic representationllmhardware designgenerative ai
Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection
The authors introduce VisAnomReasoner, a parameter-efficient vision-language model for time-series anomaly detection, addressing the lack of natural-language rationales in existing benchmarks. They construct VisAnomBench, a curated dataset with anomaly explanations selected from multiple VLMs using task-specific rewards, enabling fine-tuning for interpretable decisions. Experiments show VisAnomReasoner improves precision and F1 by ≥21.23 and 23.87 percentage points on VisAnomBench, and generalizes to TSB-AD-U with 9.57 and 13.39 percentage point gains.
vision-language modelsanomaly detectiontime-series analysisparameter-efficientinterpretable reasoning
Unlocking the Working Memory of Large Language Models for Latent Reasoning
We introduce Reasoning in Memory (RiM), a latent reasoning method that replaces autoregressive generation of intermediate tokens with fixed memory blocks, enabling compute-efficient reasoning in a single forward pass. RiM employs a two-stage curriculum: first grounding memory blocks by predicting explicit reasoning steps, then iteratively refining the final answer without step-level supervision. Experiments across language models of varying families and sizes demonstrate that RiM matches or exceeds existing latent reasoning methods while avoiding autoregressive thought generation, showing that LLMs can effectively use working memory for latent reasoning.
reasoning in memorylatent reasoningmemory blocksautoregressive generationworking memory
GPIC: A Giant Permissive Image Corpus for Visual Generation
The authors introduce GPIC, a Giant Permissive Image Corpus containing ~28 trillion pixels from 100M training, 200K validation, and 1M test images, all permissively licensed for research/commercial use. The dataset features vision-language model captions, safety filtering, deduplication, and centralized hosting on Hugging Face. They establish a benchmarking protocol for generative modeling and provide a pixel-space flow matching baseline. Resources include hosted data (Hugging Face) and an evaluation toolkit (gpic.stanford.edu).
generative modelingvision-language modelpixel-space flow matchingsafety filteringdeduplication
Locally Coherent, Globally Incoherent: Bounding Compositional Incoherence in Multi-Component LLM Agents
The paper formalizes compositional incoherence in multi-component LLM agents, where locally coherent components produce globally incoherent outputs. It introduces the compositional residual eps* to quantify this incoherence via L2 distance from the joint coherent polytope, computable at runtime. A hierarchical Boyle-Dykstra projection repairs compositions, while an e-process monitors coherence sequentially. Experiments on 1,876 ensemble cliques with four LLMs show eps* > 0 in 33-94% of cases, leading to +0.115 nats per bet regret. Three LLM-side mitigations (retrieval, partition-aware prompting, aggregator-LLM) prove ineffective or regressive.
compositional incoherencemulti-component llm agentscompositional residualboyle-dykstra projectione-process
Demystifying Data Organization for Enhanced LLM Training
The paper introduces systematic guidelines and novel methods for optimizing data organization to enhance LLM training efficiency. By reusing pre-computed sample-level scores, the authors propose four principles—Boundary Sharpening, Cyclic Scheduling, Curriculum Continuity, and Local Diversity—and implement them in two data ordering methods, STR and SAW. Experiments across various model scales and data sizes, including pre-training and SFT stages, demonstrate improved training stability and performance with minimal computational overhead.
data organizationllm trainingsample-level scorescyclic schedulinglocal diversity
Reasoning with Sampling: Cutting at Decision Points
The paper introduces Entropy-Cut Metropolis-Hastings, an algorithm that improves reasoning performance by sampling from a power distribution of a base language model. The method identifies key decision points using next-token entropy as a proxy, enabling efficient resampling at consequential reasoning steps rather than uniformly random positions. Theoretical analysis shows mixing time scales with decision points, not token count. Empirical results demonstrate consistent improvements over baselines and RL-trained models across MATH500, HumanEval, GPQA Diamond, and AIME26 benchmarks.
power distributionmetropolis-hastingsnext-token entropyreasoning tracesmixing time
RoboWits: Unexpected Challenges for Robotic Creative Problem Solving
The paper introduces RoboWits, a bi-manual robotic benchmark evaluating cognitive reasoning, creative tool use, and robustness to unexpected conditions. An automated multi-agent pipeline generates diverse tasks (30 seed tasks, 208 mutated variants) with graded difficulty across geometry, material, and assembly-based reasoning. Benchmarking reveals pre-trained vision-language-action models (VLAs) show brittleness in mutated tasks despite single-task fine-tuning success, highlighting gaps in reasoning and adaptation capabilities.
robotic benchmarkingcognitive reasoningmulti-agent pipelinevision-language-action modelstask mutation
On Language Generation in the Limit with Bounded Memory
The paper investigates language generation under bounded memory constraints, where a learner must produce new valid examples from an unknown target language while retaining limited past information. It analyzes memoryless generators, showing that countable collections of infinite languages remain generable under mild enumeration restrictions, and characterizes optimal minimax density for finite collections using combinatorial methods like Sperner's theorem. Results reveal that sliding windows of past examples do not improve worst-case density, while adaptive memory selection does. Additionally, incremental identification in the limit is shown to fail for exact identification but succeeds under approximate convergence for finite collections.
language generationbounded memoryminimax densitysperner's theoremincremental identification
In-Context Reward Adaptation for Robust Preference Modeling
We introduce In-Context Reward Adaptation, a transformer-based framework for robust preference modeling that adapts to diverse and unseen human preferences dynamically. Leveraging in-context learning, the model infers reward structures from a small set of preference demonstrations, addressing limitations of static reward models in RLHF. While standard transformers exhibit asymptotic bias to ground-truth rewards, incorporating human response time as an auxiliary signal enables successful adaptation to unseen preference domains. This approach enhances robustness in representing heterogeneous rewards and preference distribution shifts, offering a scalable solution for flexible human-AI alignment.
in-context learningreward adaptationpreference modelingtransformerhuman-ai alignment
Gram: Assessing sabotage propensities via automated alignment auditing
The paper introduces Gram, an automated framework for auditing AI alignment by assessing sabotage propensity in agentic systems. The method evaluates Gemini models across 17 simulated deployment scenarios with sabotage incentives, employing an investigator agent pipeline for targeted misbehavior analysis. Results show 2-3% misalignment rates attributed to overeagerness in role-playing and goal-seeking, with sabotage rates dropping near zero when environmental realism increases or misbehavior nudges are removed.
alignment auditingagentic systemssabotage propensityinvestigator agentovereagerness
Improved Guarantees for Heterogeneous Treatment-Effect Estimation via Matrix Completion
The paper introduces a computationally efficient estimator for heterogeneous treatment-effect estimation in panel data, framed as a matrix completion problem under low-rank assumptions. By leveraging unit-time treatment effects represented as a matrix, the method achieves row-wise ℓ₂ error bounds of Õ(√(1/n + n/m²)) without requiring propensity knowledge. The analysis provides novel perturbation bounds for low-rank approximation, advancing beyond existing spectral and Frobenius norm guarantees. Results demonstrate improved estimation accuracy for individual treatment effects compared to average-effect-focused approaches.
heterogeneous treatment effectsmatrix completionlow-rank approximationperturbation boundspanel data
Before the Shutter: Aesthetic and Actionable Portrait Photography Planning in 3D Scenes
The paper introduces 3D aesthetic portrait planning, a novel computational task for generating physically plausible and visually compelling pre-capture plans involving human pose, camera configuration, lighting, and exposure in 3D scenes. The method constructs a Photographic Scene Graph to represent scene affordances, subject-scene relations, and lighting structure, then performs aesthetic-guided comparative planning against viewfinder observations. Experiments across diverse indoor/outdoor scenes demonstrate human and MLLM evaluator preference over baselines (88.7% win rate) while maintaining photometric/geometric feasibility, shifting focus from post-production to pre-capture planning.
photographic scene graphaesthetic-guided planning3d portrait planningphotometric feasibilitypre-capture computation
Archon: A Unified Multimodal Model for Holistic Digital Human Generation
Archon introduces a unified multimodal model for holistic digital human generation, addressing the challenge of integrating text, audio, motion, and visual content. The model employs modality-specific tokenizers, a memory-efficient semantic video reparameterization for 4x token reduction, and a semantic-driven video diffusion decoder. It also features a 'Thinking in Modality' approach to decompose cross-modal tasks, enhancing fidelity and controllability. Pretrained on 72 diverse tasks, Archon demonstrates superior performance across digital human generation tasks, validated through extensive experiments.
archonmultimodaltokenizersreparameterizationdiffusion
City-Mesh3R: Simulation-Ready City-Scale 3D Mesh Reconstruction from Multi-View Images
City-Mesh3R introduces a scalable framework for reconstructing watertight 3D meshes from city-scale image collections, addressing challenges in simulation-ready geometry. The method employs topological image clustering, cluster-wise sparse SfM, and map merging, followed by geometry-aware camera selection and curvature-aware remeshing. Evaluated on city-scale datasets, it produces high-fidelity meshes with regular geometry and fine surface details, scalable to arbitrarily large scenes through distributed processing.
3d reconstructioncity-scalewatertight meshessfmremeshing
MedCase-Structured: A Text-to-FHIR Dataset for Benchmarking Diagnostic Reasoning in Clinically Realistic EHR Settings
The paper introduces MedCase-Structured, a synthetic dataset for benchmarking clinical reasoning using structured FHIR R4 bundles derived from unstructured text. A pipeline combining staged LLM generation with terminology-grounded validation ensures clinically realistic, interoperable outputs, achieving 82.5% valid FHIR generation. Evaluation shows LLMs exhibit lower diagnostic accuracy on structured FHIR inputs compared to plain text, underscoring the need for deployment-aligned clinical benchmarks.
fhirllmclinical reasoningelectronic health recordsdiagnostic accuracy
Self-Trained Verification for Training- and Test-Time Self-Improvement
The paper introduces self-trained verification (STV), a method to improve reasoning models by enhancing verification capabilities during both training and test-time. STV leverages reference solutions to train verifiers to detect self-generated errors, addressing the bottleneck in verification-refinement loops and self-training. At test time, STV significantly boosts accuracy on hard math problems (doubling it) and scientific reasoning tasks (14x improvement from 1.5% to 21%). During training, verifier-in-the-loop training (ViL) with STV feedback yields a 33% gain in pass@1 and a 30% relative improvement in standalone generator performance. This highlights the critical role of verification in advancing reasoning on complex tasks.
self-trained verificationverification-refinement loopsverifier-in-the-loop trainingreasoning modelsself-training
MIRA: Mid-training Rubric Anchoring for Source-Aware Data Selection
MIRA introduces a source-aware data selection framework for mid-training in LLM development, addressing the challenge of heterogeneous data sources with varying formats and training roles. The method employs self-anchored rubric discovery, where rubric construction is integrated into data selection: MIRA identifies evaluation criteria for each source group and distills these into scalable student scorers for corpus-wide filtering. Evaluated on code-oriented mid-training with 21 sources and 5 source groups, MIRA outperforms baseline selection methods across nine code benchmarks and achieves comparable performance to full-corpus training while utilizing only half the tokens.
mid-trainingrubric discoverysource-awaredata selectionscalable scorers
ProjectionBench: Evaluating Scientific Hypothesis Generation in LLMs Under Progressive Information Disclosure
The paper introduces ProjectionBench, a framework for evaluating scientific hypothesis generation in LLMs through progressive information disclosure. Models receive incremental technical details from research papers (topic → question → full experimental setup) and generate hypotheses at each stage, with outputs evaluated via semantic similarity to ground-truth conclusions. The method assesses both innovativeness (minimal context) and grounded reasoning (full context) across 45 materials science papers. Results show GPT-5.4 and Gemini 3.1 Pro outperform predecessors, with GPT-5.4 achieving 0.7 F1 score alignment under minimal context.
progressive information disclosuresemantic similarityhypothesis generationscientific reasoningground-truth alignment
mcp-proto-okn: Natural-language access to open scientific knowledge graphs through the Model Context Protocol
The mcp-proto-okn system introduces a Python-based Model Context Protocol server for natural-language interaction with open scientific knowledge graphs. It implements graph routing, SPARQL execution, and ontology expansion via the FastMCP framework, enabling cross-domain knowledge graph analysis. The server supports multi-graph querying and transcript generation, specifically targeting biomedical and scientific applications. Available on GitHub with full documentation, it provides a configurable client interface for AI-assisted knowledge discovery.
knowledge graphssparqlontology expansionnatural-language interfacefastmcp
Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments
Qwen-VLA introduces a unified vision-language-action foundation model for embodied intelligence, extending Qwen's capabilities to continuous action generation via a DiT-based action decoder. The model employs large-scale joint pretraining on diverse datasets (robotics trajectories, human demonstrations, synthetic data) and embodiment-aware prompt conditioning to handle multiple robot platforms. It achieves state-of-the-art performance across benchmarks: 97.9% on LIBERO, 73.7% on Simpler-WidowX, 69.0% OSR on R2R, and 76.9% OOD success in real-world ALOHA experiments, demonstrating generalization across tasks, environments, and embodiments.
embodied intelligencevision-language-actiondiffusion transformermulti-task learningout-of-distribution generalization
Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection
The paper introduces Loong, a human-like long document translation agent that addresses context window limitations and redundant information in document-level translation. Loong employs a 3E memory module (Essence-Exemplar-Entity) to store summaries, sentence pairs, and entity records, and uses observe-and-act reasoning to adaptively select optimal context. Optimized via reinforcement learning on preference data from its own trajectories, Loong achieves average gains of 13.0 points across three metrics in English ↔ Chinese, German, and French translations, demonstrating strong generalization, robustness, and stability in ultra-long documents.
document-level translation3e memory moduleobserve-and-act reasoningreinforcement learningcontext selection
LLUMI: Improving LLM Writing Assistance for Mental Health Support with Online Community Feedback
LLUMI introduces a privacy-preserving framework for mental health support using open-source LLMs, comprising a generation model (GM) for drafting responses and an improvement model (IM) for revising human-crafted responses. The system leverages Reddit community feedback, including upvotes and downvotes, to construct chosen-rejected pairs for Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO), aligning responses across readability, empathy, connection, actionability, and safety. Evaluations show LLUMI achieves performance comparable to proprietary cloud-based GPT models, demonstrating the efficacy of community-derived preference signals in training open-source models for sensitive support contexts.
supervised fine tuningdirect preference optimizationgeneration modelimprovement modelcommunity feedback
PhyGenHOI: Physically-Aware 4D Generation of Dynamic Human-Object Interactions
PhyGenHOI introduces a framework for generating physically accurate 4D Human-Object Interactions (HOI) by coupling generative human motion with object physics. The method employs a Motion Diffusion Model (MDM) for human motion and Material Point Method (MPM) for object physics, unified via 3D Gaussian Splats (3DGS). Key innovations include Windowed Attraction Loss for motion synchronization, Contact-Driven Re-simulation for momentum transfer, and Masked Video-SDS for contact fidelity. Experiments demonstrate superior performance in generating diverse, physically consistent HOI scenes compared to baselines.
4d generationhuman-object interactionmotion diffusion modelmaterial point method3d gaussian splats
How LoRA Remembers? A Parametric Memory Law for LLM Finetuning
The study introduces the Parametric Memory Law, a power law quantifying the relationship between loss reduction, effective parameters, and sequence length in Low-Rank Adaptation (LoRA) finetuning of Large Language Models. Using LoRA as a memory capacity probe, the authors identify a deterministic phase transition at the token level, where a prediction probability p > 0.5 ensures verbatim recall under greedy decoding. They propose MemFT, a threshold-guided optimization strategy that dynamically allocates training resources to sub-threshold tokens, enhancing memory fidelity and efficiency. Empirical evaluations validate MemFT's effectiveness.
parametric memory lawlow-rank adaptationphase transitiongreedy decodingmemory fidelity
Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models
The paper introduces Canonical-Context On-Policy Distillation (CCOPD) to address self-anchored drift in multi-turn language models, where partial information leads to distorted final answers. The method aligns a trainable student model processing incremental evidence with a frozen teacher conditioned on full-context prompts. Evaluated on math problems and five zero-shot task families, CCOPD improves RAW-SHARDED performance by 32% relative to the base model while maintaining full-context accuracy, demonstrating stronger evidence grounding and reduced sensitivity to earlier turn contamination.
self-anchored driftmulti-turn language modelson-policy distillationzero-shot transferevidence grounding
Reinforcement Learning with Robust Rubric Rewards
We propose Reinforcement Learning with Robust Rubric Rewards ($\text{RLR}^3$), extending Reinforcement Learning with Verifiable Rewards (RLVR) to handle multi-criteria supervision in partially verifiable vision-language tasks. $\text{RLR}^3$ routes instance-specific rubrics through two paths: LLM-as-an-extractor with deterministic verification for verifiable criteria, and LLM-as-a-Judge for non-verifiable criteria. It introduces minimal exposure to mask ground truths from extractors and images from judges, hierarchical aggregation to prioritize essential criteria, and mitigates score saturation. Evaluated on Qwen3-VL-30B-A3B across 15 benchmarks, $\text{RLR}^3$ outperforms RLVR by 4.7 points, exceeding the instruct-to-thinking model gap, and reduces exploitable false positives.
reinforcement learningrubric rewardsllm-as-an-extractorminimal exposurehierarchical aggregation
Do Language Models Track Entities Across State Changes?
The study investigates how transformer language models (LMs) perform entity tracking (ET) across state-changing operations in natural language, revealing a non-incremental mechanism. Through behavioral and mechanistic analyses of operations like PUT, REMOVE, and MOVE, the authors find LMs aggregate relevant information only at the final query token rather than tracking states incrementally. Key findings include LMs' fragile global suppression tag for REMOVE operations, which predicts observed failure modes, and a proposed mechanistic solution to nullify this tag. The work demonstrates how LM strategies for sequential tasks diverge from human-like incremental processing.
entity trackingtransformer language modelsstate-changing operationsglobal suppressionmechanistic analysis
Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning
The paper introduces GASP (Geometric-Aware Spatial Priors), a framework enhancing 3D spatial reasoning in Vision-Language Models (VLMs) by injecting geometric priors into transformer layers. GASP employs a correspondence head with deep supervision, trained via contrastive loss on point correspondences and depth consistency from video scenes, without 3D VQA data. Results show internal correspondence matching accuracy improves from <5% to >70%, with downstream gains of +18.2% on All-Angles Bench and +29.0% on VSI-Bench, demonstrating robust geometric understanding.
geometric priorsvision-language modelsspatial reasoningcontrastive lossdepth consistency
Unifying Temporal and Structural Credit Assignment in LLM-Based Multi-Agent Prompt Optimization
The paper introduces a novel method for optimizing Multi-Agent Systems (MAS) in LLM-based reasoning tasks by unifying temporal and structural credit assignment. Temporal credit uses state-space bottlenecks to pinpoint critical rounds, while structural credit employs stationary role policies to isolate agent contributions. A discrete, verbalized block coordinate descent algorithm iteratively refines prompts and aggregation protocols using LLM-generated proxy gradients. Evaluations across diverse benchmarks demonstrate reduced query complexity and improved performance, offering a principled approach to self-improving MAS.
multi-agent systemscredit assignmentblock coordinate descentproxy gradientsstate-space bottlenecks
BORA: Bridging Offline Reinforcement Learning and Online Residual Adaptation for Real-World Dexterous VLA Models
BORA introduces an offline-to-online RL framework for enhancing Vision-Language-Action (VLA) models in real-world dexterous manipulation. The method combines offline critic training with action-conditioned value guidance and online Human-in-the-Loop residual adaptation, preserving the pretrained policy as a prior while correcting execution errors. Evaluations on five dexterous tasks show 33% higher success rates than baselines and 43% improvement in unseen object generalization.
vision-language-actionoffline reinforcement learninghuman-in-the-loopdexterous manipulationresidual adaptation
When Should Models Change Their Minds? Contextual Belief Management in Large Language Models
The paper introduces Contextual Belief Management (CBM), a framework for evaluating how large language models (LLMs) manage accumulating information in long-horizon interactions. The authors propose BeliefTrack, a closed-world benchmark with Rule Discovery and Circuit Diagnosis tasks, enabling exact turn-level evaluation of belief-state alignment. Three failure modes are identified: Failed Stay, Failed Update, and Failed Isolation. While vanilla LLMs exhibit severe CBM failures and explicit prompts offer limited improvements, reinforcement learning with belief-state rewards reduces failure rates by 70.9%. Representation-level steering further decreases failures by 46.1% across tasks.
contextual belief managementbelieftrackreinforcement learningrepresentation-level steeringfailure modes
Automating Low-Risk Code Review at Meta: RADAR, Risk Calibration, and Review Efficiency
Meta's RADAR system automates code review for low-risk AI-generated changes, addressing the 105.9% YoY growth in code volume. The multi-stage funnel combines risk stratification (Diff Risk Score), static heuristics, and LLM-based review to safely approve changes. Evaluation across 535K+ diffs shows a 60.31% approve rate at 50th percentile risk threshold, with revert rates 1/3 and production incidents 1/50 of manual reviews. RADAR reduces median review time by 35% and time-to-close by 330%, demonstrating scalable risk-aware automation for AI-driven development bottlenecks.
risk stratificationdiff risk scorellm-based reviewautomated code reviewmeta radar
Persona Conditioning of Brand Recommendations in Retrieval-Augmented Commercial Chat: A Prominence-Stratified Cross-Provider Audit
The study quantifies how buyer personas condition brand recommendations in retrieval-augmented commercial chatbots, demonstrating significant persona-driven variation. Using a stratified audit design (10 personas × 8 prompts × 3 model configurations × N=10 reps), it measures Jaccard similarity drops of Δ=-0.12 to -0.20 when personas are prefixed, with effects concentrated in mid-market brands (75% recommendation swaps) versus category leaders (80% consistency). Anthropic's Claude Sonnet 4.6 shows stronger persona sensitivity (43-52% unattributed recommendations) than OpenAI models (8-29%), aligning with its retrieval-agnostic generation pathway. Results emphasize the necessity of persona-conditioned measurement in brand perception studies.
persona conditioningjaccard similarityretrieval-augmented generationprominence stratificationcontextual variation
HPO: Hysteretic Policy Optimization for Stable and Efficient Training under Sparse-Reward Regime
The paper introduces Hysteretic Policy Optimization (HPO) and its adaptive variant (A-HPO) to address sparse-reward training instability in GRPO-style reinforcement learning. HPO reduces negative-advantage update weights and replaces per-response length normalization with mean-length normalization, while A-HPO dynamically adjusts the hysteretic weight based on batch-level advantage statistics. Experiments on TeleLogs and Countdown benchmarks show A-HPO achieves 0.84 final reward (5-15% improvements over baselines) and demonstrates strongest gains in early sparse-reward regimes across 1.5B-7B models, with ablations confirming better positive/negative advantage balancing.
sparse-rewardpolicy optimizationadvantage normalizationadaptive hysteresisreinforcement learning
Double-Edged Sword or Sharp Tool? Designing and Evaluating Triadic LLM-Teacher Collaboration for K-12 Writing at Scale
This paper contributes a triadic LLM-teacher-student collaboration system for K-12 writing education, evaluated through a multidimensional framework based on Systemic Functional Linguistics and suggestion trajectory tracing. The study analyzes 57,954 essays from 10,195 students across 120 schools over two years, demonstrating improved writing quality via strategic labor division: LLMs as generative engines reduce teacher burnout, while teachers ensure feedback quality as pedagogical gatekeepers. Results reveal a ceiling effect where excessive linguistic expansion yields diminishing returns, suggesting adaptive collaboration is needed as student proficiency increases.
triadic collaborationsystemic functional linguisticssuggestion trajectory tracinggenerative enginepedagogical gatekeeper
What drives performance in molecular MPNNs? An operator-level factorial benchmark
The study introduces an operator-level factorial benchmark to analyze performance drivers in 2D molecular message-passing neural networks (MPNNs), decomposing them into message-seed initialization, node-edge fusion, and node update operators. Evaluating 84 configurations across ten MoleculeNet datasets under controlled conditions, the analysis reveals that message construction primarily influences performance, with message-seed initialization showing significant family-level effects for both regression and classification. Concatenation-based node-edge fusion outperforms Hadamard gating in differentiating chemically distinct heteroatoms and resisting oversmoothing. Selected configurations achieve competitive performance, ranking best on eight datasets. The findings offer empirical design heuristics for molecular MPNNs by focusing on chemical information flow in the message-passing pipeline.
message-passing neural networksmoleculenetnode-edge fusionoversmoothingfactorial benchmark
Token-Level Generalization in LoRA Adapter Backdoors: Attack Characterization and Behavioral Detection
The study demonstrates that LoRA adapters in fine-tuned LLMs can be reliably backdoored via training data poisoning while maintaining baseline task performance. Using a Qwen 2.5 1.5B prompt-injection classifier, the authors show that token-level generalization in backdoors (e.g., activating on RFC references but not structurally similar citations) favors attackers. They characterize the attack across model scale, LoRA rank, and trigger strings, and propose two detection methods: a behavioral detector (using outlier_gap and mean_attack_rate) and a weight-level statistic (cross-module Frobenius norm deviation). Both methods achieve perfect separation of poisoned adapters, with causal patching identifying the MLP block as the backdoor location. The behavioral detector transfers across models without retuning.
lora adaptersbackdoor attacktoken-level generalizationfrobenius normbehavioral detection
CalArena: A Large-Scale Post-Hoc Calibration Benchmark
We introduce CalArena, a large-scale benchmark for evaluating post-hoc calibration methods across diverse machine learning tasks and models. The benchmark aggregates predictions from classical models, deep learning architectures, and foundation models, providing unified implementations of dozens of calibration methods within a reproducible framework. We propose Post-Hoc Improvement (PHI) in proper scoring rules as a principled evaluation metric, capturing both calibration quality and predictive performance. Our comprehensive empirical study reveals that smooth calibration functions outperform binning-based approaches, dedicated multiclass methods are essential in high-dimensional settings, and generic machine learning models require calibration-specific design. We release all data, code, and evaluation tools to facilitate future research.
post-hoc calibrationproper scoring rulesmulticlass classificationdeep learningfoundation models
Modularizing Educational LLM-Agency for Fostering Responsible Learning Assistance
The paper proposes a modular agentic AI chatbot architecture for responsible educational assistance, addressing structural limitations of monolithic LLM-based systems. The architecture incorporates pedagogical principles through specialized modules tailored to different stages of exercise solving, aiming to mitigate risks such as reduced transfer capabilities and impaired critical thinking. The authors identify desiderata for responsible LLM deployment in education and argue for modularization to enhance controllability, transparency, and oversight in learning processes. The proposed design enables targeted pedagogical guidance while maintaining adaptability to diverse educational contexts.
agentic architecturepedagogical principlesexercise solvingmonolithic llmtransfer capabilities
iLoRA: Bayesian Low-Rank Adaptation with Latent Interaction Graphs for Microbiome Diagnosis
The paper introduces iLoRA, a Bayesian graph-conditioned LoRA framework for parameter-efficient adaptation of LLMs, specifically addressing microbiome diagnosis. iLoRA dynamically infers latent interaction graphs from input data to generate input-conditioned LoRA updates, jointly learning prediction and interaction structure. Evaluated on interactive QA with human-annotated graphs and multi-cohort IBD diagnosis, iLoRA outperforms LoRA and Bayesian baselines, recovers graphs aligned with human annotations, and provides calibrated uncertainty with moderate computational overhead.
bayesianlow-rank adaptationlatent interaction graphsmicrobiome diagnosisparameter-efficient
Dissociative Identity: Language Model Agents Lack Grounding for Reputation Mechanisms
The article critiques the applicability of human reputation mechanisms to autonomous language model agents, arguing that their dissociative nature undermines key trust properties. Analyzing agents as mutable module assemblages (foundational models, system prompts, tool-access policies), the authors demonstrate how dissociativity prevents identifiability, predictability, and sanction internalization. They propose replacing identity-based governance with protocol-based behavioral harnesses to address this structural mismatch.
language model agentsreputation mechanismsdissociative identitybehavioral continuityprotocol-based governance
BioRefusalAudit: Auditing Biosecurity Refusal Depth Using General and Domain-Fine-Tuned Sparse Autoencoders
This paper introduces BioRefusalAudit, a method to audit biosecurity refusal depth in language models using general and domain-fine-tuned sparse autoencoders (SAEs). The study evaluates five architectures, revealing inconsistent refusal behaviors: Gemma 2 2B-IT hedged on all hazard-adjacent queries, while Gemma 4 E2B-IT refused 65/75 prompts only with chat-template formatting. Qwen 2.5 1.5B and Phi-3-mini over-refused benign biology (83-87%). A divergence score D compared surface responses to internal SAE activations, showing a 0.647-point gap between comply and refuse responses in Gemma 4. The study highlights activation-level auditing as a tool to uncover failure modes missed by behavioral evaluation.
biosecuritysparse autoencodersrefusal depthactivation auditingdivergence score
On Distributional Reinforcement Learning in Chaotic Dynamical Systems
The paper demonstrates that distributional reinforcement learning (RL) offers improved learning stability in chaotic dynamical systems, where traditional RL methods suffer from high-variance bootstrap targets due to exponential sensitivity to initial conditions. By analyzing return distributions under the $1$-Wasserstein metric, the authors show these evolve more regularly than individual trajectories, yielding a smoother distributional Bellman objective. This provides a principled explanation for distributional RL's advantages in chaotic environments, linking optimization geometry to system dynamics.
distributional reinforcement learningchaotic dynamical systemswasserstein metricbellman objectivegradient conditioning
Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents
The paper introduces Metacognitive Memory Policy Optimization (MMPO), a novel method for optimizing memory policies in long-horizon LLM agents by focusing on belief clarity rather than trajectory-level success. MMPO employs Belief Entropy, a self-supervised proxy measuring epistemic uncertainty about the latent task state, to penalize ambiguous recursive summaries that degrade memory quality. Experiments demonstrate MMPO's superiority over existing methods, maintaining 97.1% performance in 1.75M-token contexts across diverse long-horizon tasks.
memory-augmented llm agentsbelief entropymetacognitive optimizationlong-horizon reasoningepistemic uncertainty
Neural Network Verification using Partial Multi-Neuron Relaxation
The paper introduces partial multi-neuron relaxation, a novel approach for neural network verification that balances tightness and scalability. By generating multi-neuron bounds only for a heuristically selected subset of neurons, the method improves upon single-neuron relaxation's loose bounds while avoiding the computational cost of full multi-neuron relaxation. The technique leverages existing branching heuristics for neuron selection and bounding hyper-plane optimization, implemented within the Marabou verifier. Experimental results demonstrate superior performance compared to existing bound tightening methods, highlighting its potential for efficient formal verification of deep neural networks.
neural network verificationpartial multi-neuron relaxationbound tighteningmarabou verifierbranching heuristics
Do Proactive Agents Really Need an LLM to Decide When to Wake and What to Anchor?
The paper introduces a temporal-graph-learning (TGL) model as an efficient alternative to LLM-based proactive agents for processing structured event streams. By treating user activity as graph updates rather than text, the TGL encoder computes per-event trigger probabilities and routing scores in a single forward pass, invoking LLMs only when necessary. Evaluated across 14 backbones, TGL improves mean F1 by +16.7 (up to +46.0), achieves superior trigger AUCs, and operates at 11.13 ms/event on GPU (13.99 ms on laptop), with a 220 MiB BF16 footprint enabling on-device deployment.
temporal-graph-learningllmtrigger probabilitystructured event streamauc
Temporal Stability and Few-Shot Prompting in Math Task Assessment
This longitudinal study investigates temporal stability and few-shot prompting effects on AI tools' ability to classify mathematics tasks using the Task Analysis Guide (TAG). It evaluates Gemini (general-purpose) and Coteach (education-specific) models across baseline, version updates, and few-shot prompting conditions with two exemplars per cognitive demand category. Results show version updates yielded mixed effects: Gemini maintained 58% accuracy while Coteach decreased from 75% to 50%. Few-shot prompting improved both models: Gemini increased to 67% and Coteach recovered to 75%. Findings suggest prompt engineering reliably outperforms passive model updates for specialized educational tasks, informing AI tool selection and implementation strategies.
task analysis guidefew-shot promptingtemporal stabilitycognitive demandprompt engineering
Anchorless Diversification for Parallel LLM Ideation
The study introduces anchorless methods for diversifying candidate-idea pools generated by large language models (LLMs) in creative tasks, comparing them to anchor-based approaches. It evaluates independent generation and semantic direction stratification against self-, peer-, and representative-anchor baselines across three creative task families, under neutral and population-referential divergent instructions. Results show that population-referential divergence is a strong low-cost baseline for increasing semantic diversity while preserving quality proxies. Semantic direction stratification outperforms others, organizing generations across broad semantic directions with a single planning call, achieving the best diversity-quality-compute frontier. Anchored regeneration shows strength in final-pool diversity but loses advantage under full-pipeline token accounting.
semantic direction stratificationpopulation-referential divergenceanchorless methodsparallel inferencetoken accounting
Overcoming Forgetting in LLM Fine-Tuning with Evolution Strategies
This paper demonstrates that Evolution Strategies (ES) fine-tuning induces performance drift rather than irreversible forgetting in large language models (LLMs), with prior-task performance often recovering during training. It identifies that such drift arises from ES training dynamics, particularly random walk behavior in weakly constrained weight-space directions. To mitigate this, the authors propose Anchored Weight Decay (AWD), a parameter-space regularization technique that constrains optimization toward initial model parameters. AWD effectively stabilizes prior-task performance while preserving target-task performance, achieving benefits comparable to large ES population sizes at reduced computational cost. The findings position ES as a viable approach for continual learning in LLMs.
evolution strategieslarge language modelsanchored weight decayparameter-space regularizationcontinual learning
AgentSchool: An LLM-Powered Multi-Agent Simulation for Education
AgentSchool introduces an LLM-powered multi-agent simulator for education that models learning as state transitions rather than prompted role-play. The system features cognitively growable student agents with knowledge graphs, workflow pools, and misconceptions, paired with adaptive teacher agents that operate within the Zone of Proximal Development. Experiments demonstrate differentiated mastery traces, ZPD-consistent adaptation patterns, and plausible social dynamics like clique formation, outperforming baseline simulators in capturing educational complexity.
multi-agent simulationknowledge graphszone of proximal developmentllm-powered agentseducational ai
Enhancing Multi-Agent Communication through Attention Steering with Context Relevance
The paper introduces Agent-Radar, a training-free context management method for LLM-based multi-agent systems that addresses performance degradation from long conversation histories. The method dynamically steers agent attention using temporal and spatial decay mechanisms to maintain focus on relevant context. Experiments across five benchmarks show absolute performance gains up to 7.64 points, with robustness demonstrated under increasing agent counts and interaction rounds. Ablation studies confirm the necessity of core components.
multi-agent systemscontext managementattention steeringtemporal decayspatial decay
DAMEL: Dual-Axis Multi-Expert Learning for Class-Imbalanced Learning
The paper introduces DAMEL, a dual-axis multi-expert learning algorithm for class-imbalanced learning. DAMEL reduces prediction bias and variance by employing multiple experts along representation and time axes: it concatenates expert representations for auxiliary balanced classifier training and aggregates network weights across epochs. Experiments show DAMEL effectively mitigates both bias and variance in imbalanced datasets.
class-imbalanced learningmulti-expert learningrepresentation axistime axisprediction variance
PARCEL: Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding
PARCEL introduces a novel visual tokenization architecture for efficient vision-language understanding by dynamically partitioning feature extraction. The method combines spatial pool tokens as low-frequency layout anchors with conditioned elastic query tokens through Pool-Conditioned Query Resampling, reducing redundancy in spatial mapping. Evaluated across 27 benchmarks, PARCEL outperforms existing matryoshka baselines in performance-efficiency trade-offs while maintaining the 'train once, deploy anywhere' paradigm.
vision-language modelstoken compressionpool-conditioned resamplingspatial groundingefficient inference
Beyond MSE: Improving Precipitation Nowcasting with Multi-Quantile Regression
This study demonstrates that multi-quantile regression improves precipitation nowcasting performance compared to standard pointwise losses. The authors reformulate training as a multi-quantile regression problem using SmaAt-UNet, comparing MSE, MAE, and multi-quantile pinball-loss on radar precipitation data from the Netherlands. Results show an 8.6% reduction in test-set MSE versus MSE-trained models, while upper-quantile outputs enable better risk-sensitive prediction of heavy rainfall. The approach requires no architectural changes or generative sampling, offering a simple alternative to pointwise loss optimization. Implementation is publicly available on GitHub.
precipitation nowcastingmulti-quantile regressionsmaat-unetpinball-lossrisk-sensitive prediction
No More K-means:Single-Stage Sparse Coding for Efficient Multi-Vector Retrieval
The paper introduces Single-stage Sparse Retrieval (SSR), a novel multi-vector retrieval method that replaces computationally expensive K-means clustering with sparse coding via Sparse Autoencoder (SAE). SSR projects token embeddings into high-dimensional sparse representations, enabling efficient inverted indexing while preserving semantic granularity. Evaluated on the BEIR benchmark, SSR achieves 15x faster indexing, 50% lower retrieval latency, and improved accuracy compared to ColBERTv2.
multi-vector retrievalsparse autoencoderinverted indexingtoken embeddingsbeir benchmark
Evolving Features vs Evolving Entire Trees with GP for Interpretable Survival Analysis
The study introduces a genetic programming approach to enhance both accuracy and interpretability in survival analysis by evolving feature sets and jointly optimizing survival tree structures. It contrasts evolving individual features against full tree evolution, focusing on shallow trees that require expressive feature combinations. Experiments on two real-world datasets demonstrate that evolutionary feature construction improves predictive performance across various tree induction strategies, with joint evolution yielding the highest potential for producing interpretable, high-performing shallow survival trees.
genetic programmingsurvival analysistree inductionfeature constructioninterpretability
VLA-Trace: Diagnosing Vision-Language-Action Models through Representation and Behavior Tracing
VLA-Trace introduces a diagnostic framework for analyzing Vision-Language-Action (VLA) models through representation dynamics, causal control attribution, and behavioral manifestation. The method combines cross-modal and checkpoint-drift centered kernel alignment (CKA) to trace representation evolution, attention knockout interventions to identify modality-specific control pathways, and rollout-level behavioral probes to examine grounding, shortcut dependence, and semantic following. Experiments on π₀.₅ and OpenVLA reveal distinct modality-specific adaptation dynamics, different multimodal routing strategies, and limitations in fine-grained semantic following despite strong visually grounded trajectory generation. These findings suggest future directions for representation-preserving adaptation, causal VLA circuits, and compositional semantic control.
kernel alignmentattention knockoutmodality-specificsemantic followingrepresentation dynamics
xModel-KD: Cross-modal Knowledge Distillation for 3D Scene Perception using LiDAR
The paper proposes xModel-KD, a cross-modal knowledge distillation framework for 3D point cloud segmentation that addresses annotation scarcity and modality limitations. The method learns unified per-point representations through a cross-modal fusion encoder with contrastive learning, aligning 2D texture features from images with 3D geometric features from LiDAR. By integrating pre-trained backbones, it transfers appearance cues to geometry-aware point features. Experiments demonstrate a 2% mIoU improvement over LiDAR-only baselines, showing effective multi-modal representation learning for annotation-efficient 3D scene understanding.
knowledge distillation3d segmentationcross-modal learningcontrastive learningpoint clouds
When Cloud Agents Meet Device Agents: Lessons from Hybrid Multi-Agent Systems
This work systematically examines the design space of hybrid multi-agent systems (MASs) combining on-device small language models (SLMs) and cloud-based large language models (LLMs) for AI inference. Two representative MAS architectures were adapted to support hybrid inference, analyzing how design choices impact the Pareto frontier of power, cost, and performance. Results reveal that while SLMs benefit from LLM assistance, optimal architectures are highly task-dependent, and increased compute does not consistently improve performance, highlighting the nuanced trade-offs in hybrid MAS design.
multi-agent systemssmall language modelslarge language modelshybrid inferencepareto frontier
How Reliable Are AI Attackers Against a Fixed Vulnerable Target? A 400-Run Empirical Study of LLM Penetration Testing Consistency
This work provides the first large-scale empirical measurement of LLM attack consistency through 400 autonomous penetration testing runs (4 models, 100 each) against a fixed OWASP Juice Shop honeypot. Holding prompt, orchestrator, and target constant, models exhibited statistically significant (p < 0.001) differences in exploitation rates: Gemini 2.5 Flash-Lite (85/100), Claude Sonnet 4 (61/100), GPT-4o-mini (56/100), and qwen2.5-coder:14b (25/100). Failure modes were model-specific, including API truncation (Claude), premature completion (qwen), and iteration exhaustion (GPT-4o-mini). First-exploit timing consistently fell within 15-30 seconds, with cross-service credential reuse observed only in models retaining extensive conversation history.
llm penetration testingautonomous cyber attacksexploitation consistencyhoneypot evaluationapi truncation
PokerSkill: LLMs Can Play Expert-Level Poker without Training or Solvers
PokerSkill introduces a training-free, solver-free framework that combines rule-based poker skills with LLMs to achieve expert-level play in imperfect-information games. The method uses a deterministic context engine to retrieve relevant skill fragments from a human-designed library, constraining LLM actions to reasonable choices. Evaluated against GTOWizard, GPT-4-turbo achieves -57 mbb/hand and Claude Opus 4.7 achieves -87 mbb/hand, reducing losses by 49-61% versus baselines and outperforming Slumbot, demonstrating competitive performance without game-specific training or solver access.
pokerlarge language modelsimperfect-information gamesrule-based skillsaction-grounding
Selective QA over Conflicting Multi-Source Personal Memory: A Diagnostic Testbed and Method Comparison
The paper introduces a diagnostic benchmark for selective question answering (QA) over conflicting multi-source personal memory, addressing the challenge of resolving inconsistencies in persistent AI agents. The benchmark comprises 18 question templates across 8 reasoning types, with 34,560 instances featuring controlled distortions and deterministic ground truth. Evaluations compare baselines (no/single-source access), structured fusion methods, and large language models (LLMs), with the best fusion resolver achieving 80.3% accuracy (85.3% selective accuracy with abstention) and the top LLM reaching 70.0% accuracy (71.0% selective accuracy). The dataset and generation process are publicly released.
selective qamulti-source memoryconflict resolutiondiagnostic benchmarkabstention
Conformal Certification of Reasoning Trace Prefixes
The paper introduces CROP (Conformal Reasoning Output Prefixes), a method for certifying valid prefixes in language model reasoning traces while controlling error probability. CROP uses step-level risk proxies to identify the longest contiguous prefix below a calibrated threshold, routing uncertified suffixes for review. Evaluated on six process-labeled reasoning datasets, results show that standard metrics like AUROC inadequately capture prefix utility, whereas certified prefix length better evaluates verifiers. CROP improves downstream repair accuracy by preserving valid intermediate steps and discarding erroneous suffixes.
conformal predictionreasoning tracesprefix certificationrisk proxyprocess supervision
A Predictive Law for On-Policy Self-Distillation From World Feedback
The paper identifies a linear correlation between the initial performance gap and final improvement in on-policy self-distillation (OPSD), providing a predictive law for anticipating training outcomes without full execution. This relationship holds across context types and model families, demonstrating scalability with model size and suggesting implications for in-context learning. Results indicate OPSD performance can be preemptively predicted and tuned, offering a principled approach to integrating world feedback in post-training pipelines.
on-policy self-distillationworld feedbackpredictive lawin-context learningscaling laws
Projectional Decoding: Towards Semantic-Aware LLM Generation
The paper introduces projectional decoding, a framework for improving semantic validity in LLM-generated software artifacts by maintaining a partial graph model alongside text during generation. This dual representation enables incremental semantic validation, explicit uncertainty modeling, and error detection while guiding generation toward provably correct outputs. Preliminary results on program generation demonstrate improved semantic validity. The approach aims to enable verifiable automation across software engineering tasks by bridging LLM outputs with domain-specific reasoning.
projectional decodingsemantic validationpartial graph modelllm generationsoftware artifacts
REPOT: Recoverable Program-of-Thought via Checkpoint Repair
RePoT (Recoverable Program-of-Thought) introduces a deterministic verified replay mechanism for Program-of-Thought (PoT) by identifying the first invalid transition in a Python-generated action plan and resuming execution via a single LLM call. This approach incurs at most one additional LLM call on ~14% of failing PoT instances. RePoT outperforms PoT by +3 to +11 percentage points across four closed-model configurations on PuzzleZoo-775, achieving 96.9% accuracy on gpt-5.4-mini-medium versus PoT's 86.3%. It also demonstrates significant gains on Gemini (+3.8pp) and PlanBench Blocksworld (+1.1 to +11.4pp), with checkpoint information proving critical for recovery, as evidenced by >=30% success on GPT-medium and >=70% on Gemini in Derail-550.
recoverable program-of-thoughtdeterministic verified replaycheckpoint repairllm callverified-prefix
Robust and Generalizable Safety Steering for Text-to-Image Diffusion Transformers
SafeDIG introduces a robust safety steering framework for Diffusion Transformers (DiTs) in text-to-image generation, addressing the challenge of harmful semantics propagating through layered cross-modal processes. The method employs position-aware sparse feature transfer, constructing Sparse Autoencoders (SAEs) at distinct intervention positions and using robustness-aware pre-training routing to prioritize stable sites. It freezes the SAE encoder as a reusable sparse safety dictionary while adapting the decoder to target-domain activation manifolds. During inference, SafeDIG combines Blend and Repel operations to steer unsafe activations. Experiments on FLUX.1 Dev and Stable Diffusion 3.5 Large demonstrate reduced unsafe generation rates while maintaining source-domain safety and image quality.
diffusion transformerssparse autoencodersposition-awaresafety steeringactivation manifold
Masked Diffusion Modeling for Anomaly Detection
The paper proposes Masked Diffusion for Anomaly Detection (MaskDiff-AD), a forward-only method using masked diffusion models trained on nominal data to detect anomalies in categorical, mixed-type, and discrete sequence data. The approach constructs anomaly scores by measuring reconstruction difficulty of randomly masked coordinates, avoiding reverse-time sampling while maintaining content sensitivity. Evaluated on fourteen tabular datasets (ADBench, UADAD) and four text datasets (NLP-ADBench), MaskDiff-AD achieves competitive performance, ranking first overall against twelve tabular baselines with theoretical guarantees on Type-I/II errors.
masked diffusionanomaly detectiondiscrete sequencesnominal datatype-i/ii errors
Learning to Choose: An Empowerment-Guided Multi-Agent System with semantic communication for Adaptive Method Selection
The paper introduces a multi-agent framework combining contextual bandits, structured inter-agent communication, and semantic checkpoints to prevent semantic drift in automated scientific workflows. Building on the ATHENA framework and empowerment theory, the system integrates specialized LLM agents, grounded code generation, and self-healing execution loops. Experiments on sensitivity analysis and uncertainty quantification workflows demonstrate improved convergence (38% reduction in semantic drift), robustness, and adaptation to novel contexts compared to baseline methods.
semantic driftcontextual banditsmulti-agent systemsempowerment theoryself-healing execution
Token Inflation: How Dishonest Providers Can Overcharge for Large Language Model Usage
The paper identifies a trust paradox in per-token billing for commercial LLMs, where providers can manipulate token counts due to unverifiable auditing frameworks. Analyzing three recent token auditing methods, the authors demonstrate systematic inflation possibilities: hidden reasoning usage can be inflated by 1,469% undetected, while tokenization ambiguity permits 50.85% over-reporting even with visible reasoning strings. Results indicate fundamental flaws in provider-controlled evidence chains, suggesting solutions like trusted execution attestation or cryptographic proofs for honest billing.
token inflationper-token billingtrust paradoxauditing frameworkstokenization ambiguity
Domain-Specific Data Synthesis for LLMs via Minimal Sufficient Representation Learning
The paper introduces DOMINO, a framework for domain-specific data synthesis that learns minimal sufficient representations from reference examples without requiring explicit domain descriptions. The method combines prompt tuning with contrastive disentanglement to separate domain-level patterns from noise, theoretically expanding synthetic data diversity. Evaluated on coding benchmarks with implicit domain definitions, DOMINO improves Pass@1 accuracy by up to 4.63% over instruction-tuned baselines, demonstrating robust domain adaptation without manual prompt engineering.
domain-specific data synthesisminimal sufficient representationcontrastive disentanglementprompt tuninginstruction-tuned backbones
Alignment-Guided Score Matching for Text-to-Image Alignment in Diffusion Models
The paper introduces Alignment-Guided Score Matching (AGSM), a reward-free post-training method that improves text-to-image alignment in diffusion models by integrating contrastive guidance directly into the score-matching objective. Unlike prior contrastive approaches like SoftREPA that suffer from over-penalization of negative pairs (manifesting as over-counting/repetition), AGSM refines soft text tokens via score-level alignment directions, yielding more semantically faithful generations. Experiments demonstrate a 35% improvement in counting accuracy on GenEval while maintaining SoftREPA's performance, with compatibility across SD1.5, SDXL, and SD3 backbones and complementarity to RL-based methods.
text-to-image alignmentdiffusion modelsscore matchingcontrastive learningsoft tokens
Teaching Values to Machines: Simulating Human-Like Behavior in LLMs
This work demonstrates that value-prompted Large Language Models (LLMs) can exhibit human-like value structures and value-behavior relationships when assessed using psychological questionnaires. The authors conducted large-scale experiments (5M+ questions) comparing leading LLMs to human baselines, employing validated psychological instruments from value theory. Results show strong alignment between value-induced LLMs and humans at both individual and population levels, with human value distributions improving simulation fidelity. The study establishes LLMs as psychologically grounded tools for human behavior simulation.
large language modelsvalue theorypsychological questionnaireshuman behavior simulationvalue-behavior alignment
Audio Jailbreaks in Large Audio-Language Models: Taxonomy, Attack-Defense Analysis, and Cost-Aware Evaluation
This paper introduces a unified taxonomy and empirical evaluation framework for jailbreak attacks and defenses in Large Audio Language Models (LALMs), addressing heterogeneous threat models and evaluation protocols. The taxonomy categorizes attacks into semantic, acoustic, signal, and embedding-layer types, and defenses into guard-based, training-free, and training-based approaches. Controlled experiments across ten open-source LALMs measure attack success rate, benign refusal, and latency. Key findings include Acoustic Best-of-N exposing severe audio-space vulnerabilities, Narrative Framing as a low-latency semantic threat, and defenses compromising robustness for usability. The study advocates for cost- and utility-aware evaluation in LALM safety benchmarks.
jailbreak attacksaudio-language modelstaxonomyempirical evaluationlatency
RAISE: RAG Design as an Architecture Search Problem
The paper introduces RAISE, a framework for treating Retrieval-Augmented Generation (RAG) design as an architecture search problem, enabling systematic evaluation of hyperparameter optimization methods. RAISE implements 13 search algorithms and evaluates them across seven text and multimodal datasets with three random seeds, providing standardized search spaces and budgets. Results indicate task-dependent optimization performance, cautioning against universal strategy rankings and advocating for reproducible RAG research.
retrieval-augmented generationhyperparameter optimizationarchitecture searchmultimodal datasetsreproducible research
Give it Space! Explicit Disentangling of Positional and Semantic Representations in Encoders
The paper proposes explicitly disentangling positional and semantic representations in Transformer encoders by processing three separate streams: semantic, absolute positional (AP), and relative positional (RP). This modification confines masked-language-modeling (MLM) to the semantic stream, enabling mechanistic analysis. Key findings include: (1) AP collapses into a low-frequency 2D manifold capturing document structure; (2) attention heads specialize into structure/semantic groups with RP supporting semantics; (3) standard positional encodings poorly retain macroscopic structure. The approach improves linguistic representation on 49/65 phenomena in Flash-Holmes probing.
positional encodingtransformer encodersemantic representationattention headsprobing benchmark
Test Time Training for Supervised Causal Learning
The paper introduces Test-Time Training for Supervised Causal Learning (TTT-SCL), addressing three limitations of existing supervised causal learning methods: poor out-of-distribution generalization, fragility to distribution shifts, and compositional generalization failures. TTT-SCL dynamically generates training sets aligned with each test instance, leveraging connections to score-based methods and an efficient training-set generation module. Experiments on synthetic, pseudo-real, and real-world datasets show TTT-SCL outperforms both traditional SCL and causal discovery baselines.
supervised causal learningtest-time trainingcausal discoverydistribution shiftcompositional generalization
From GPS Points to Travel Patterns: Flexible and Semantic Trajectory Generation with LLMs
The paper proposes HTP, a hierarchical framework for generating realistic urban trajectories using LLMs. It first quantizes GPS trajectories into travel pattern tokens via a residual quantization VAE, capturing spatial irregularities, then extends LLM vocabulary with these tokens for conditional generation. Supervised fine-tuning enables flexible trajectory synthesis under varying conditions. Evaluations on real-world datasets show HTP outperforms baselines by 29.78% in generation quality.
trajectory generationresidual quantization vaetravel pattern tokensllm fine-tuningurban dynamics
VisualThink-VLA: Visual Intermediate Reasoning for Effective and Low-Latency Vision-Language-Action Policies
VISUALTHINK-VLA introduces a visual intermediate-reasoning framework for vision-language-action (VLA) policies, addressing latency and interference issues in textual chain-of-thought approaches. The method employs a compact visual-evidence interface and selective routing mechanism to preserve spatial precision and enable low-latency inference. Evaluated on benchmarks including BridgeData V2, it achieves a 22.8× speedup (0.367s vs. 8.377s) and higher success rates compared to reasoning-augmented baselines, while maintaining real-time closed-loop execution.
visual intermediate-reasoningvision-language-action policiesselective routing mechanismlatency reductionvisualevidence-set
Discovering Cooperative Pipelines: Autoresearch for Sequential Social Dilemmas
The paper introduces a two-level autoresearch framework where an outer-loop AI agent (researcher $\mathcal{R}$) autonomously redesigns the inner-loop pipeline of an LLM policy-synthesis system for Sequential Social Dilemmas (SSDs). The researcher edits code, prompts, and evaluation logic, optimizing for utilitarian efficiency or Rawlsian maximin welfare. Results show the framework outperforms hand-designed baselines, reduces variance, and adapts pipelines to objectives (e.g., injecting fairness mechanisms under maximin). Evaluations span two games (Cleanup, Gathering) and two policy-synthesizer LLMs, demonstrating objective-dependent pipeline discovery.
autoresearchsequential social dilemmaspolicy-synthesiswelfare objectivesfairness mechanisms
KairosAgent: Agentic Time Series Forecasting with Fused Semantic Reasoning
KairosAgent introduces a novel agentic framework for multimodal time series forecasting, addressing limitations in semantic reasoning and numerical comprehension. The framework integrates an LLM-based reasoner and a TSFM-based forecaster, dynamically invoking analytical tools to enhance LLMs' numerical understanding and semantic reasoning. Reasoning results are fused into the TSFM pipeline, improving prediction accuracy. A large-scale corpus of high-quality trajectories and a reinforcement learning from forecasting paradigm with multi-turn refinement and turn-level credit assignment further enhance reasoning. Experiments show KairosAgent achieves superior zero-shot forecasting performance, maximizing the utility of pretrained LLMs and TSFMs.
multimodal time seriessemantic reasoningllm-based reasonertsfm-based forecasterzero-shot forecasting
Cookie-Bench: Continuous On-screen Key Interaction Evaluation for Web Generation
The paper introduces Cookie-Bench, a 1,000-query WebDev benchmark spanning 11 domains and 54 leaf tasks, balanced across difficulty tiers and target-language groups, designed to evaluate interactive web generation. The proposed evaluation framework employs metacognitive monitoring across three stages: Static Perception, Agent-Driven Interaction (capturing screen video/audio), and Dynamic Scoring for holistic functionality/aesthetics assessment. Results show close alignment with human ratings while revealing performance gaps across 13 frontier LLMs on interactive tasks.
web generationreference-free evaluationmetacognitive monitoringagent-driven interactiondynamic scoring
Accelerating Constrained Decoding with Token Space Compression
CFGzip introduces token space compression to accelerate constrained decoding in LLMs, reducing overhead from context-free grammar (CFG) compliance. The method compresses the token search space offline, enabling efficient selection of CFG-conforming tokens during generation. Experiments demonstrate latency reductions of up to two orders of magnitude and a 7.5x speedup in total constrained generation time when integrated with a state-of-the-art grammar engine. This advancement makes complex CFG-constrained decoding feasible at scale.
context-free grammarconstrained decodingtoken space compressionlatency reductionoffline optimization
Genetically Aligned Patient Representations Improve Hematological Diagnosis
The study introduces a genetically aligned patient representation framework that enhances hematological diagnosis by multimodal alignment of single white blood cell images with karyotype and somatic mutation data. The method employs a two-stage approach: (1) self-supervised pretraining of a transformer aggregator using an iBOT head on 1500+ patients, followed by (2) genetic alignment via supervised contrastive loss for acute myeloid leukemia. Results show improved diagnostic performance over slide-level histopathology foundation models and enable disease-specific retrieval. The framework aligns with clinical workflows and is publicly available.
multimodal alignmenttransformer aggregatoribot headsupervised contrastive losshematological diagnosis
Evaluating Skill and Stability of ArchesWeather and ArchesWeatherGen under Multi-Decadal Climate Simulations
The study evaluates ArchesWeather and ArchesWeatherGen, originally weather forecasting models, as forced atmospheric models for climate simulation under AIMIP Phase 1 protocol. By conditioning on monthly mean SST and SIC, the models demonstrate stable long-term climate simulations, accurate annual cycles, and fidelity to ERA5 climatology. Results show robust performance in reproducing large-scale circulations, interannual variability, and distribution tails, comparable to numerical climate models.
archesweatherarchesweathergenaimipsstsic
Compass: Navigating Global Marine Lead Data Integration through Expert-Guided LLM Agent
The authors present Compass, an expert-guided LLM agent framework for extracting marine lead (Pb) data from unstructured academic papers without fine-tuning. The method employs a Knowledge Tree co-designed with domain experts to decompose tasks into verifiable steps, ensuring scientific validity. Applied to 230,000 papers, Compass extracted 3,751 new Pb records with 92% accuracy, creating the largest integrated marine Pb database and improving coverage in under-sampled regions like the East China Sea.
llm agentknowledge treemarine leaddata extractionscientific validity
Meta-Programming for Linear-time Temporal Answer Set Programming
The authors present a meta-programming framework for implementing temporal extensions of Answer Set Programming (ASP), addressing the rigidity of existing ASP systems. The approach extends clingo's theory grammar with type specifications and nesting capabilities, while introducing a transformation pipeline to preserve nested modalities during grounding. The framework demonstrates extensibility through meta-encodings for temporal equilibrium logic (TEL), metric temporal logic (MEL), and dynamic equilibrium logic (DEL), with specific handling of interval constraints and Fischer-Ladner closure. The resulting metasp system encapsulates this workflow.
answer set programmingtemporal equilibrium logicmeta-programmingclingofischer-ladner closure
Honeyval: A Comprehensive Evaluation Framework for LLM-powered HTTP Honeypots
The paper introduces Honeyval, a unified evaluation framework for LLM-powered HTTP honeypots, addressing scalability, reproducibility, and practical attack representation gaps in existing methods. The framework grounds evaluations in 16 backend applications, employs AI hacking agents as attackers, and defines verifiable exploit goals. Experiments demonstrate LLM-powered honeypots achieve longer attacker interactions (2.5× rule-based baselines) with lower detection rates (30% reduction), while maintaining cost efficiency. Trade-offs between interaction duration and detection risk are quantified across counter-offensive configurations.
honeypotsllm-poweredhttpevaluation frameworkai hacking agents
Hijacking Agent Memory: Stealthy Trojan Attacks Through Conversational Interaction
The paper introduces MemPoison, a novel memory poisoning attack targeting LLM agents' long-term memory through dialogue interactions. The attack bypasses selective memory mechanisms via three components: semantic relational bridges binding triggers/payloads, entity masquerading to resist rewriting, and joint embedding optimization for stealth. Evaluations show attack success rates up to 0.95 across domains, exploiting embedding-space anisotropy and attention pattern shifts. Defenses exhibit fundamental limitations against the attack.
memory poisoningllm agentsembedding-space anisotropyselective memorytrigger-payload binding
Formalizing Mathematics at Scale
AutoformBot introduces a multi-agent system for large-scale mathematical autoformalization, converting informal textbook content into verified Lean 4 code. The system employs thousands of LLM agents with formal verification tools, dependency-aware scheduling, and collaborative version control to process 26 open-access textbooks across analysis, algebra, topology, combinatorics, and probability. The resulting Atlas library contains 45,000 Lean 4 declarations and 500K lines of code, demonstrating the feasibility of graduate-level mathematics autoformalization. The authors release both AutoformBot and Atlas as open-source artifacts.
autoformalizationlean 4multi-agent systemformal verificationmathematical library
MuPHI: Learning Implicit Multimodal Harm Reasoning via Semantically Grounded Reward Optimization
The paper introduces MuPHI, a dataset for evaluating multimodal harm detection through implicit reasoning, and proposes MuPHIRM, a training framework for vision-language models (VLMs). MuPHI contains image-text pairs with annotated harm rationales across diverse categories, requiring context-dependent reasoning beyond surface features. MuPHIRM optimizes multi-perspective rewards to jointly learn harm semantics and reasoning chains, improving detection accuracy and out-of-distribution robustness over baselines. Results demonstrate that reward-driven reasoning enhances generalization beyond benchmark-specific shortcuts in VLMs.
multimodal reasoningharm detectionreward optimizationvision-language modelsout-of-distribution robustness
HoliTok:A Coutinuous Holistic Tokenization with Robust Dual Capabilities of Speech Generation and Understanding
The paper introduces HoliTok, a continuous holistic tokenization model for unified speech generation and understanding. HoliTok encodes 48kHz speech into 25Hz sequences of 128-dimensional latents using a progressive training strategy that preserves signal fidelity and semantic information. The model supports both AR+DiT architectures for synthesis and recognition tasks without additional optimization. Experiments demonstrate competitive reconstruction fidelity, improved generative learnability, and robust performance in unified generation-understanding tasks. Results indicate HoliTok's effectiveness as a foundational representation interface for spoken language modeling.
holistic tokenizationspeech generationsignal fidelityprogressive trainingunified modeling
Make LLM Learn to Synthesize from Streaming Experiences through Feedback
The paper introduces StreamSynth, a novel setting where synthetic data generation tasks arrive sequentially, enabling experience accumulation across tasks. It proposes SynLearner, a framework that learns reusable synthesis patterns from feedback, balancing quality and diversity as tasks evolve. Experiments across multiple benchmarks demonstrate SynLearner's ability to transfer knowledge from earlier tasks, improving performance on subsequent ones by 12-18% compared to isolated synthesis approaches.
streamsynthsynlearnersynthetic data generationexperience accumulationcross-task transferability
CityGen: Structure-Guided City-Style Synthesis for Cross-City Autonomous Driving
We propose CityGen, a diffusion-based generative framework for zero-label city adaptation in autonomous driving, addressing domain shifts in appearance, road topology, and traffic patterns across cities. CityGen synthesizes city-style data via HD-map-conditioned generation guided by city-level visual prompts, eliminating the need for labeled target data or city-specific annotations. Evaluated on CityTransfer-Bench, a geographically disjoint benchmark for cross-city generalization, CityGen demonstrates consistent improvements in robustness across perception, segmentation, and planning tasks, establishing a scalable and label-efficient approach for generalizable autonomous driving systems.
diffusion-based generationhd-map-conditioned synthesiszero-label adaptationcross-city generalizationcitytransfer-bench
It`s All About Speed: AI`s Impact on Workflow in Music Production
The study investigates AI's impact on music production workflows through ethnographic analysis of professional recording engineers, mixers, and producers. Researchers examined adoption patterns of automated tools and identified key tensions regarding workflow efficiency, controllability, and creative agency. Findings reveal tool design considerations to mitigate conflicts between speed optimization and artistic control in AI-assisted music production environments.
workflow optimizationautomated mixingcreative agencycontrollabilityethnographic study
Toward AI Systems That Understand Self and Others: A Multi-Phase Inference Framework for Human Cognitive Diversity and World-Model Alignment
The paper proposes a Multi-Phase Inference Mechanism (MIM) to formalize how heterogeneous world models emerge from cognitive diversity, addressing AI systems' capacity to understand divergent human reasoning. MIM introduces a phase-formation space, foregrounding field, subject-specific profile states, and alignment maps to model variations in inferential targets, state representations, and update priorities. This framework redefines world-model alignment as enabling mutual processability of heterogeneous representations rather than enforcing consensus, with applications to philosophical disagreements, cognitive typology, and AI alignment. The approach aims to make differences in meaning, value, and prediction error computationally tractable for improved human-AI interaction.
multi-phase inference mechanismworld-model alignmentphase-formation spaceforegrounding fieldalignment maps
Label Over Logic? How Source Cues Bias Human Fallacy Judgments More Than LLMs
This study investigates source-label bias in fallacy detection, comparing human and LLM reasoning. Using an online experiment (N=505) with five source conditions (human, AI, human with AI assistance, AI with human assistance, no disclosure), participants evaluated comments containing logical fallacies alongside LLMs (GPT-5.2, Gemini 2.5 Flash, Claude Sonnet 4.5). Results show humans significantly favored fallacies labeled as human-written or human-assisted, assigning higher trust and evaluation ratings, while LLMs maintained stable evaluations across source labels. Confidence levels remained high for both groups regardless of fallacy presence, suggesting source-label bias is primarily a human vulnerability.
source-label biaslogical fallaciesllm evaluationreasoning qualityhuman-ai collaboration
Does The Way You Plan Matter? An Empirical Study of Planning Representations for LLM Web Agents
The paper introduces PlanAhead, a static planner-executor framework investigating how natural language plan representations affect LLM-based web agents. It categorizes WebArena tasks into 3 difficulty levels automatically and evaluates 4 plan representations (sequential subgoals, narrative, pseudocode, checklist) on hard tasks using multimodal LLMs (OpenAI, Alibaba, Google). Two novel metrics, Achievement Rate (AR) and Solved-Task Consistency (STC), reveal that both plan formulation and underlying LLM significantly impact agent robustness and task success.
plan representationweb agentsmultimodal llmsachievement ratesolved-task consistency
On the Geometry of Games and their Solvers
The paper introduces a framework for understanding equilibrium computation in games through a solver-game map, linking games to solver dynamics via learned structure-aware representations. By formalizing solver synthesis with a structure recognizer and adaptive policy, it identifies continuous regions of algorithmic validity and overlapping solver behavior. Empirical results demonstrate that fixed primitives suffer regime mismatch, while learned representations organize game space into solver-aligned clusters, revealing a structured geometry of solvability.
equilibrium computationsolver-game mapstructure-aware synthesisalgorithmic validitygeometry of games
Selection Hyper-heuristics Can Automatically Adjust the Learning Period to Optimally Solve Pseudo-Boolean Problems
The paper presents an automated method for determining the optimal learning period τ in the Random Gradient hyper-heuristic when applied to pseudo-Boolean optimization problems. By dynamically adjusting τ, the algorithm eliminates manual parameter tuning while maintaining theoretical guarantees. The authors prove that their approach selects the optimal neighborhood size in a 1-o(1) fraction of iterations, achieving near-optimal runtime for the LeadingOnes benchmark. This advancement improves upon traditional hyper-heuristics by incorporating adaptive learning periods rather than relying solely on immediate iteration feedback.
hyper-heuristicspseudo-booleanleadingonesrandomised local searchneighborhood size
Agora: Toward Autonomous Bug Detection in Production-Level Consensus Protocols with LLM Agents
We introduce Agora, a domain-aware multi-agent framework for detecting protocol-level logic bugs in consensus implementations, addressing limitations of LLM-based approaches in analyzing complex state-dependent behaviors. Agora integrates hypothesis-driven testing with LLM capabilities through specialized agents that collaboratively explore protocol state spaces, synthesize attack scenarios using domain-specific constraints, and validate findings via iterative refinement. Evaluated on four consensus protocols (Raft, EPaxos, HotStuff, BullShark) with four state-of-the-art LLMs, Agora discovers 15 previously unknown protocol-level logic bugs violating safety properties, outperforming existing LLM-based agents which detect zero such bugs.
consensus protocolsmulti-agent frameworkprotocol-level logic bugshypothesis-driven testingstate-dependent behaviors
Redundant or Necessary? A Benchmark for Detecting Redundant Steps in Agent Trajectories
The paper introduces RedundancyBench, a benchmark for detecting redundant steps in LLM-based agent trajectories, addressing the overlooked aspect of execution efficiency in current evaluations. The benchmark comprises diverse tasks with annotated trajectories where each step is labeled for its task contribution. Three detection methods are evaluated, with the best achieving only 24.88% F1 score, underscoring the task's difficulty and the need for further research. Code and dataset are publicly available.
llm-based agentsexecution efficiencyredundant step detectiontask trajectoriesbenchmark evaluation
Internal Representation, Not Clinical Knowledge: Where Apparent LLM Triage Failures Originate
The study investigates why large language models (LLMs) exhibit high under-triage rates in clinical-triage benchmarks when constrained to multiple-choice output formats, despite performing better with free-text responses. Using sparse-autoencoder (SAE) features in Gemma 3 4B/12B IT and Qwen3-8B, the authors find that medical features remain consistent across formats but are suppressed during multiple-choice decisions. Three methods—natural-language autoencoder verbalization, decision-token logit attribution, and top-feature characterization—reveal that scaffold and format features, not medical knowledge, drive decision logits. Behavioral experiments confirm the gap stems from output format misalignment (e.g., off-by-one errors) rather than clinical representation failures.
sparse-autoencoderlogit attributionunder-triageclinical representationoutput format
LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training
The paper introduces LaRA, a layer-wise representation analysis framework for detecting data contamination in RL post-trained LLMs, addressing limitations of output-level detection methods. LaRA employs three geometric metrics—perturbation sensitivity, directional collapse, and local representation rigidity—to identify progressive deviations across model layers caused by contamination. Experiments demonstrate that LaRA's aggregation of layer-wise representation metrics outperforms existing output-level baselines in contamination detection for RL-trained reasoning models.
data contaminationreinforcement learninglayer-wise analysisrepresentation geometryperturbation sensitivity
CRITIC-R1: Learning Structured Critics for Retrieval-Augmented Generation
CRITIC-R1 introduces a structured critic framework for retrieval-augmented generation (RAG) that addresses hallucinations and reasoning errors by formulating critique as an explicit error diagnosis problem. The method categorizes RAG errors into diagnostic dimensions and employs reinforcement learning with two reward functions: Conservative Judgement Alignment (CJA) for calibrated high-level judgements and Diagnostic Quality Alignment (DQA) for fine-grained feedback. Training utilizes GRPO-based RL with supervision from external LLM teachers. Experiments on five QA benchmarks demonstrate consistent improvements in answer quality over strong RAG baselines.
retrieval-augmented generationreinforcement learningerror diagnosisconservative judgement alignmentdiagnostic quality alignment
Mitigating Hallucination in Vision-Language Models through Barrier-Regulated Adaptive Closed-form Steering
BRACS (Barrier-Regulated Adaptive Closed-form Steering) mitigates object hallucination in large vision-language models (LVLMs) by adaptively steering hidden states only when visual grounding deteriorates. The method monitors attention-based grounding and computes corrective updates in closed form, requiring no auxiliary training. Evaluated on LLaVA-1.5-7B and Qwen-VL-Chat, BRACS reduces CHAIR$_s$ by 9.4 points and improves POPE F1 by 2.7 points while maintaining performance on general multimodal benchmarks, operating at 80% of greedy decoding throughput.
vision-language modelsobject hallucinationclosed-form steeringvisual groundingadaptive correction
Evolutionary Dynamics of Cooperation in Next-Generation LLM Agent Systems: A Cross-Provider Empirical Extension
This study extends evolutionary game theory benchmarks for cooperative behavior in LLM agents to four frontier models (Claude Sonnet 4.6, Gemini 2.5 Flash, Gemini 3.1 Pro, GPT-5.4 Mini) released in 2025-2026. Using the Iterated Prisoner's Dilemma (IPD) framework, the authors evaluate three prompting styles (Default, Prose, Self-Refine) and four population compositions (balanced, biased, with and without noise). Results show persistent cooperative biases across providers (9/12 model-prompt combinations favor cooperative equilibria), substantial cross-provider divergence (Gemini 2.5 Flash reaches 77% aggressive equilibria), and partial support for aggressive capability parity. Provider identity emerges as the strongest correlate of equilibrium outcomes, while noise remains a universal challenge.
iterated prisoner's dilemmacooperative biasprompting stylespopulation compositionsequilibrium outcomes
Moment-KV: Momentum-Based Decode-Time KV Cache Compression for Long Generation
Moment-KV introduces a momentum-based KV cache compression method for LLM decoding phases, addressing limitations of static heuristics in long-generation tasks. By modeling token importance as a temporally evolving state with decayed attention aggregation, it captures both sustained long-term relevance and short-term bursts. Experiments demonstrate 2.3-3.2% generation fidelity improvements while preserving decoding latency, outperforming rigid recency-based approaches.
kv-cacheattention dynamicsdecode-time compressionmomentum aggregationlong-generation
Mitigating Stethoscope-Induced Shortcuts in Respiratory Sound Classification under Federated Domain Generalization with Causality-Inspired Interventions
A causality-inspired federated domain generalization (FedDG) framework is proposed to mitigate stethoscope-induced shortcuts in respiratory sound classification (RSC). The method integrates a causality-inspired device style intervention network for content-preserving style perturbations, counterfactual text augmentation to neutralize metadata shortcuts, and gradient alignment for device-invariant representations across clients. Built on a multimodal language-audio pretraining model, the framework outperforms conventional data augmentation and federated learning baselines in leave-one-device-out validation on the ICBHI and SPRSound datasets, addressing inter-stethoscope variability in multi-site deployment.
federated domain generalizationrespiratory sound classificationcounterfactual augmentationgradient alignmentmultimodal pretraining
Towards Verifiable Multimodal Deep Research: A Multi-Agent Harness for Interleaved Report Generation
The authors propose \textsc{Ptah}, a multi-agent system for verifiable multimodal deep research that generates interleaved text-image reports. The framework employs specialized agents for visual-aware planning, evidence collection with a Visual Working Memory, and declarative multimodal composition, supervised by a verifier agent enforcing factual grounding and cross-modal consistency. Evaluated via the new \textsc{Ptah}Eval protocol, the system outperforms baselines in reliability, visual informativeness, and usability on deep research benchmarks.
multimodal deep researchvisual working memorymulti-agent systemdeclarative tool usecross-modal consistency
ESPO: Early-Stopping Proximal Policy Optimization
ESPO (Early-Stopping Proximal Policy Optimization) improves reinforcement learning for large language models by terminating rollouts upon detecting reasoning failures. The method computes surrogate regret using existing logits during sampling, truncates trajectories when cumulative regret exceeds estimates, and treats failures as absorbing states with terminal rewards. Evaluated on DeepSeek-R1-Distill-Qwen-7B for mathematical reasoning, ESPO outperforms PPO on AIME~2024 (46.28% vs. 45.25%), AMC~2023 (85.83% vs. 82.94%), and MATH-500 (87.42% vs. 85.43%), while reducing rollout tokens by over 20%.
early-stoppingproximal policy optimizationsurrogate regretabsorbing statestemporal-difference errors
HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM Quantization
HARP introduces a learnable structured two-sided orthogonal processor for extreme low-bit LLM quantization, replacing fixed Hadamard transforms with adaptive rotations. The method employs sparse butterfly-like block-orthogonal stages, supports non-power-of-two dimensions via Mixed-Radix schedules, and initializes to randomized Hadamard transforms. Evaluated on models from 1B to 70B parameters at 2-4 bit settings, HARP improves perplexity and zero-shot accuracy over fixed rotations while maintaining deployment efficiency (128 tok/s vs. 61 tok/s for FP16).
post-training quantizationhadamard transformlow-bit quantizationsparse butterflymixed-radix
CB-SLICE: Concept-Based Interpretable Error Slice Discovery
CB-SLICE introduces a concept-based error slice discovery method leveraging Concept Bottleneck Models (CBMs) to identify systematic model failures tied to human-interpretable concepts. By grouping samples with shared concept prediction errors and pinpointing responsible keyword concepts, it provides fine-grained explanations directly linked to inference failures. Evaluations across multiple benchmarks demonstrate CB-SLICE outperforms state-of-the-art methods in bias detection while offering more faithful error explanations.
concept bottleneck modelserror slice discoverymodel debuggingsystematic errorsinterpretable explanations
OmniMatBench: A Human-Calibrated Multimodal Reasoning Benchmark Across 19 Materials Science Subfields
The authors introduce OmniMatBench, a human-calibrated multimodal reasoning benchmark spanning 19 materials science subfields, addressing the gap in evaluating AI systems' ability to reason from materials knowledge to applications. The benchmark comprises 3,171 expert-curated QA and calculation problems across four domains: fundamental knowledge, structural/engineering materials, processing/manufacturing, and functional/applied materials. Evaluation of 13 MLLMs reveals a maximum score of 0.372, demonstrating significant limitations in materials-science reasoning, with observed deficiencies in subfield consistency, reasoning heuristics, knowledge distribution, and high-level application under formula/retrieval/code assistance.
multimodal reasoningmaterials sciencebenchmark evaluationexpert-curatedknowledge application
OptSkills: Learning Generalizable Optimization Skills from Problem Archetypes via Cluster-Based Distillation
OptSkills introduces an archetype-centric skill learning system for optimization problems, improving generalization by clustering problems by underlying archetypes rather than surface narratives. The method distills successful optimization trajectories into reusable workflow-level skills within each cluster, refining or expanding the skill library for out-of-distribution generalization. The system achieves 68.27% micro-averaged accuracy on diverse datasets, 26.91% on MIPLIB-NL (outperforming DeepSeek-V3.2-Thinking by 4.53%), and 72.79% on OOD NLCO after skill learning on Nano-CO.
optimization skillsarchetype clusteringworkflow distillationout-of-distribution generalizationllm-based optimization
Towards Localized and Disentangled Knowledge Editing for Multimodal Large Language Models
We propose Localized and Disentangled Knowledge Editing (LDKE), a framework addressing Causal Misalignment and Feature Entanglement in Multimodal Knowledge Editing for MLLMs. LDKE introduces a Fast Localization module to identify critical layers and a Disentanglement Classifier to route inputs, enabling precise edits while preserving unrelated knowledge. Experiments across benchmarks demonstrate LDKE's superior performance in propagating edits to related contexts with high locality, outperforming existing methods in maintaining accuracy and generalization.
multimodal knowledge editingcausal misalignmentfeature entanglementdisentanglement classifierfast localization
Quantifying and Optimizing Simplicity via Polynomial Representations
The paper introduces polynomial representations as a quantitative measure of neural network simplicity, demonstrating their predictive value for generalization. The method approximates network behavior along data-dependent paths using orthogonal polynomial bases, yielding low-dimensional functional representations whose effective degree serves as a simplicity metric. Experiments show this metric outperforms existing proxies like sharpness across tasks and architectures, while the derived differentiable simplicity regularizer improves generalization in image/text classification, vision-language fine-tuning, and reinforcement learning.
polynomial representationssimplicity biasgeneralization metricsorthogonal basesdifferentiable regularizer
Inferring Code Correctness from Specification
The paper introduces TRAILS~, a method for validating LLM-generated code correctness by grounding reasoning in concrete input-output pairs rather than direct code analysis. It generates diverse test inputs via category partitioning from specifications, executes them against candidate code, and uses LLMs to assess specification conformance. Evaluated on LiveCodeBench and CoCoClaNeL with Qwen3Coder-30B, Devstral-Small-24B, and Olmo3.1-Instruct, TRAILS~ improves Matthew Correlation Coefficient by up to 39% over Zero-Shot COT and outperforms HoarePrompt, while demonstrating greater stability across seeded runs and broader code coverage.
llm-generated codespecification conformancecategory partitioningmatthew correlation coefficientinput-output pairs
Harnessing non-adversarial robustness in large language models
The work proposes debiasing for robustness, a fine-tuning method to enhance Large Language Model (LLM) robustness against semantically-neutral prompt variations without full retraining. Theoretical analysis identifies perturbation-induced bias in neural modules as a key robustness factor. Experiments demonstrate that debiasing improves robustness and provides certification against random prompt perturbations, with conditions identified for its effectiveness. The approach offers a computationally efficient alternative to adversarial training.
large language modelsrobustnessfine-tuningperturbation-induced biasprompt variations
PRAIB: Peer Review AI Benchmark of Behaviour of LLM-Assisted Reviewing
The paper introduces Peer Review AI Benchmark (PRAIB), a framework for evaluating LLM-generated peer reviews against human norms using metrics for specificity, style, and engagement behavior. The authors analyze 11,000 reviews from five LLMs for 1,000 ICLR and NeurIPS papers (2021–2025), comparing them to human feedback across prompting strategies. Results show LLM reviews are less variable, positively biased, overconfident, and model-dependent in cross-referencing, while being longer and more complex but often missing atomic weaknesses identified by humans.
peer reviewlarge language modelsbenchmarkingbehavioral divergenceprompting strategies
Data filtering methods for training language models
The study compares Confident Learning and Dataset Cartography for label error detection across three Russian text classification corpora (ru_emotion_e-culture, RuCoLA, TERRa) using a fine-tuned rubert-base-cased model. Results indicate method effectiveness depends on dataset characteristics: Confident Learning improves F1-macro significantly on small, noisy datasets, while Dataset Cartography removes fewer examples conservatively. Both methods outperform random removal, validating their utility for targeted data filtering.
confident learningdataset cartographylabel error detectiontext classificationrubert
AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security
The paper introduces AgentDoG 1.5, a lightweight and scalable alignment framework addressing safety risks in open-world AI agents like OpenClaw and Codex. The method updates safety taxonomies for emergent risks, employs a taxonomy-guided data engine with influence-function purification, and trains parameter-efficient variants (0.8B-8B) using only ~1k samples. Results show performance parity with closed-source models (e.g., GPT-5.4), 100x Docker-level deployment efficiency gains, and state-of-the-art safety moderation in interactive agentic scenarios. All models and datasets are open-sourced.
agent alignmentinfluence-function purificationsafety taxonomyparameter-efficientonline guardrail
SAAS: Self-Aware Reinforcement Learning for Over-Search Mitigation in Agentic Search
The paper introduces SAAS, a reinforcement learning framework addressing over-search in agentic search systems by cultivating dynamic self-awareness. SAAS employs (i) search boundary modeling through contrasting search-disabled/enabled rollouts, (ii) boundary-aware rewards penalizing unnecessary searches, and (iii) stage-wise optimization prioritizing reasoning before search regularization. Experiments show SAAS significantly reduces over-search while preserving accuracy, with code released anonymously.
agentic searchover-search mitigationreinforcement learningsearch boundary modelingstage-wise optimization
SkillsInjector: Dynamic Skill Context Construction for LLM Agents
SkillsInjector introduces a dynamic skill injection method for LLM agents, addressing limitations of static skill selection through a two-stage adaptive approach. A context planner learns execution-grounded preferences and adaptively selects skills, while a set-aware renderer tailors skill descriptions relative to co-injected neighbors. Evaluated on tau2-bench, SkillsBench, and ALFWorld, SkillsInjector improves over baselines by 3.9, 6.1, and 7.3 percentage points respectively, with ablations confirming contributions from selection, budgeting, and rendering components.
llm agentsskill injectioncontext planningset-aware renderingadaptive budgeting
MEMENTO: Leveraging Web as a Learning Signal for Low-Data Domains
MEMENTO introduces a framework leveraging the web as a learning signal for low-data domains, contrasting with approaches relying on labeled or pseudo-labeled data. It employs an Adaptive Exploration Tree (AET) for iterative web exploration within sessions, decomposing tasks into evolving questions and reflecting on intermediate findings. Across sessions, it accumulates experience via dual-channel memory, separating declarative knowledge from procedural knowledge. Evaluated on sales automation and legal research, MEMENTO outperforms ReAct baselines by +25.6% and +36.5%, respectively, demonstrating scalable acquisition of task-specific expertise.
adaptive exploration treedual-channel memorylow-data domainsweb explorationtask-specific expertise
Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems
We propose Meta-Team, a collaborative self-evolution framework for LLM-based multi-agent systems (MAS) that addresses challenges in experience-driven evolution. Meta-Team preserves agent execution contexts, coordinates post-task communication, and enables multi-scale self-evolution by transforming execution experience into reusable improvements across agent behaviors, inter-agent coordination, and team-level organization. Evaluated on six long-horizon agent benchmarks, Meta-Team outperforms single-agent systems, hand-crafted MAS, and prior MAS evolution methods, demonstrating enhanced reliability and scalability in MAS self-evolution.
multi-agent systemsself-evolutionexecution contextlong-horizon taskscollaborative learning
Certified Policy Optimisation for Nested Causal Bandits via PAC-Bayes Risk
The paper introduces Nested Causal Thompson Sampling (NCTS) for Nested Contextual Causal Bandits (NCCBs), a hierarchical structural causal model where actions at one level influence context distributions at subsequent levels. The method employs mechanism-factorised belief sampling per episode and recursive action selection, with a theoretical contribution of a causal PAC-Bayesian excess-risk bound for off-policy certification. Experiments demonstrate NCTS's superior zero-shot transfer under distribution shifts compared to RFF-GP joint regression, with recursive meta-to-inner commitment outperforming joint-commit alternatives. The results enable progressive certified handover for safe deployment across timescales.
nested causal banditspac-bayes riskthompson samplingstructural causal modelzero-shot transfer
Croissant Tasks: A Metadata Format for Reproducible Machine Learning Evaluations
The authors introduce Croissant Tasks, a declarative metadata format for machine learning evaluations that enables conceptual reproducibility by decoupling task specifications from implementations. The method combines (1) a formal specification abstracting low-level details, (2) an automated LLM pipeline for retrofitting existing benchmarks, and (3) autonomous agent-based reproduction pipelines. Empirical validation demonstrates that agents can generate functional reproduction pipelines from Croissant specifications, addressing scalability limitations of manual reproducibility efforts.
conceptual reproducibilitymetadata formatautonomous agentsllm pipelinetask specification
Hista and Numca: Estimate State Value Effectively for LLM Reinforcement Learning
We introduce Hista and Numca, two novel techniques for improving state value estimation in reinforcement learning (RL) for large language models (LLMs). Numca utilizes numerical spans as gradable milestones, while Hista employs LLM hidden states to weight-average disjoint rollouts and their returns. Both methods are evaluated on the State Value Estimation Benchmark (SVEB), demonstrating that they outperform standard approaches like PPO, which collapse to coarse group-average baselines. Experiments show that Hista and Numca yield more accurate state value estimates and enhance training performance across various RL algorithms and model sizes, without significant computational overhead.
state value estimationreinforcement learninglarge language modelsppohidden states
Energy-Aware NECO for Single-Pass Pixel-wise Out-of-Distribution Detection in Semantic Segmentation
The paper proposes Energy-Aware NECO, a single-pass pixel-wise out-of-distribution (OOD) detector for semantic segmentation that combines a centered NECO-style geometric ratio with a logit-based Energy score. The method standardizes both components using in-distribution validation statistics and fuses them via convex combination. Evaluated on miniMUAD with pixel-level OOD labels, the hybrid score achieves 0.8539 AUROC, outperforming NECO-only (0.8280), Energy-only (0.8171), and a predictive-entropy baseline (0.8124), while maintaining single-pass efficiency.
semantic segmentationout-of-distribution detectionenergy scoresingle-pass inferenceuncertainty estimation
From XXLTraffic to EvoXXLTraffic: Scaling Traffic Forecasting to Sensor-Evolving Networks
The authors introduce XXLTraffic and EvoXXLTraffic, a dataset family for traffic forecasting that addresses the limitation of fixed sensor sets in existing benchmarks. XXLTraffic spans 27 years of California PeMS and Transport for NSW data, supporting long-term forecasting with multi-year gaps. EvoXXLTraffic reorganizes this data to include per-year active sensors, traffic-flow matrices, and graph snapshots, with sensor growth ratios ranging from +305% to over +10,000%. A yearly streaming forecasting protocol is defined, and various baselines, including spatio-temporal GNNs and evolving-graph continual methods, are benchmarked. Results show that many state-of-the-art methods fail on this ultra-large evolutionary dataset, highlighting its realism and utility.
traffic forecastingspatio-temporal gnnevolving-graphstreaming protocolsensor-evolving networks
LFQ: Logit-aware Final-block Quantization for Boosting the Generation Quality of Low-Bit Quantized LLMs
The paper introduces Logit-aware Final-block Quantization (LFQ), a method to enhance block-wise post-training quantization (PTQ) for large language models by addressing generative quality degradation. LFQ quantizes the final Transformer block using cross-entropy minimization between full-precision and quantized model logits, aligning token probability distributions. This approach improves generation accuracy over state-of-the-art PTQ while maintaining performance on language modeling and understanding tasks.
quantizationlogit-awaretransformercross-entropygeneration
Benchmarking Positional Encoding Strategies for Transformer-Based EEG Foundation Models
This study benchmarks five positional encoding strategies for transformer-based EEG foundation models, addressing the challenge of generalizing across EEG tasks. Using the CBraMod backbone, the authors evaluate Spherical Positional Encoding (SPE) and Asymmetric Conditional Positional Encoding (ACPE) under linear probing and fine-tuning protocols for motor imagery classification and emotion recognition. Results indicate task-dependent performance: SPE excels in motor imagery but underperforms in emotion recognition, while ACPE shows more consistent cross-task performance. The findings highlight the absence of a universal positional encoding solution for EEG decoding scenarios.
positional encodingeeg foundation modelsspherical positional encodingasymmetric conditional positional encodingmotor imagery classification
A unified deeplearning framework for contrast-phase-specific virtual monochromatic imaging
A unified deep learning framework synthesizes contrast-phase-specific virtual monochromatic 50 keV images from single-energy CT (SECT) data by leveraging contrast phase information as a prior. The model integrates contrast phase priors into the energy transformation process using a novel prior conditioning architecture, trained on DECT-derived 70 keV and 50 keV image pairs across four contrast phases: Angio, Arterial, Portal, and Delayed. Results demonstrate effective contrast enhancement and generalization across phases, enabling the generation of 50 keV-like images from SECT inputs while preserving phase-specific dynamics.
virtual monochromatic imagingsingle-energy ctcontrast phaseprior conditioningenergy transformation
Why Specialist Models Still Matter: A Heterogeneous Multi-Agent Paradigm for Medical Artificial Intelligence
The paper proposes HetMedAgent, a heterogeneous multi-agent framework for medical AI that orchestrates collaboration between generalist LLMs, domain-specific specialist models, and clinicians. The system features conflict-aware evidence fusion, uncertainty-based clinician intervention triggering, and adaptive threshold calibration. Experiments on three clinical decision-making tasks show that combining generalist LLMs with specialist models outperforms using either type alone, demonstrating the continued importance of domain-specific models for modality-specific analysis. The approach shifts focus from monolithic medical foundation models to multi-agent collaboration.
heterogeneous multi-agentmedical artificial intelligenceconflict-aware fusionadaptive threshold calibrationmodality-specific analysis
Citation-Closure Retrieval and Per-Rule Attribution for Real-World Regulatory Compliance Question Answering
The paper introduces RefWalk, a framework for regulatory compliance QA that addresses citation-closure retrieval and per-rule attribution challenges. It formalizes the task with RegOps-Bench, a benchmark featuring an Operational Knowledge Graph derived from complex regulations. RefWalk traverses cross-document citations, aggregates multi-view candidates, and enforces explicit source mapping. Evaluations on a HIPAA dataset show improved retrieval recall and citation accuracy, highlighting limitations of existing systems in handling flat-structure rules.
regulatory compliancecitation-closureknowledge graphretrieval recallattribution
Multi-Legal-Bench: Evaluating LLMs on Legal Reasoning Across Jurisdictions, Languages, and Legal Traditions
We introduce Multi-Legal-Bench, the first cross-jurisdictional legal benchmark evaluating identical tasks across six countries, four language families, and 134 million court decisions. The benchmark defines five tasks—court-type classification, judgment form classification, case-outcome prediction, legal norm extraction, and cause category prediction—mapped to structured metadata from national court registries, forming a sparse 5x6 task-jurisdiction matrix. We evaluate seven frontier LLMs under zero-shot and 3-shot prompting via AWS Bedrock, with additional scaling analysis on smaller models (3-12B). Results show task-dependent few-shot effects replicate across jurisdictions, no single model dominates language rankings, cross-lingual transfer quality is better predicted by label-set alignment than language proximity, and tokenizer fertility does not significantly predict cross-lingual accuracy.
cross-jurisdictionalfew-shot promptingtask-jurisdiction matrixlabel-set alignmenttokenizer fertility
Uncertainty-Aware Transfer Learning for Cross-Building Energy Forecasting: Toward Robust and Scalable District-Level Energy Management
The paper proposes an uncertainty-aware transfer learning framework for cross-building energy forecasting using Temporal Fusion Transformer (TFT), evaluated on high-resolution sub-meter data from Aalborg University (source) and NEST building (target). Key innovations include the Transfer Robustness Index (TRI) for quantifying generalization, and a layer-freezing ablation showing Probe-Only fine-tuning (455/806K parameters updated) achieves best transfer quality (TRI=3,097). Monte Carlo Dropout yields 93.2% prediction interval coverage, while data-scarcity analysis demonstrates monotonic improvement with target-domain data.
temporal fusion transformertransfer robustness indexmonte carlo dropoutcross-building forecastinglayer-freezing ablation
NaRA: Noise-Aware LoRA for Parameter-Efficient Fine-Tuning of Diffusion LLMs
The paper proposes Noise-aware Low-Rank Adaptation (NaRA), a parameter-efficient fine-tuning method for Diffusion Large Language Models (dLLMs) that addresses the limitations of static PEFT approaches like LoRA. NaRA introduces a hypernetwork-conditioned low-rank core matrix that dynamically adapts to noise levels during the diffusion process, maintaining computational efficiency while capturing trajectory-dependent variations. Experiments demonstrate consistent improvements over noise-agnostic baselines on commonsense reasoning, mathematical reasoning, and code generation tasks.
diffusion llmsparameter-efficient fine-tuninglow-rank adaptationhypernetworknoise-aware
The Little Book of Generative AI Foundations: An Intuitive Mathematical Primer
The book offers a mathematically rigorous yet accessible introduction to the theoretical foundations of generative AI, focusing on derivations and conceptual connections rather than implementation specifics. It systematically traces the lineage of key generative model families, including probabilistic PCA, variational autoencoders, diffusion models, normalising flows, autoregressive models, GANs, and energy-based approaches. By maintaining mathematical precision while emphasizing structural relationships, the primer aims to equip researchers and students with foundational understanding for advanced study or application of generative modeling techniques.
variational autoencodersnormalising flowswasserstein gansdiffusion modelsenergy-based models
Teaching Language Models to Check Grounded Claim Factuality with Human Test-Taking Strategies
We propose a method for grounded claim factuality checking by formulating it as a true/false reading comprehension task and prompting large language models (LLMs) with explicit test-taking strategies. This approach reduces token usage by over 80% compared to unguided reasoning and achieves state-of-the-art performance on one benchmark. To further reduce inference costs, we train small language models (SLMs) via supervised fine-tuning and a self-revision mechanism, enabling them to perform competitively with strong baselines while generating interpretable rationales. The method demonstrates efficiency and accuracy across two factuality benchmarks.
grounded claim factualitytest-taking strategiessupervised fine-tuningself-revision mechanisminterpretable rationales
Personalized Turn-Level User Conversation Satisfaction Benchmark
The paper introduces PersTurnBench, a benchmark for personalized turn-level user conversation satisfaction evaluation, addressing the limitation of existing methods that measure generic response quality. The proposed evaluator combines compact user memories with target-turn context to generate satisfaction scores and dissatisfaction-oriented rationales. Meta-evaluation shows improvements in ordinal agreement (22.5% over baselines) and dissatisfied-turn detection (15.3% F1 gain) through personalized memory and score calibration, enabling model comparison without new human labels.
personalized evaluationturn-level satisfactionuser memoryscore calibrationconversational ai
BitTP: The Lightweight Trajectory Prediction Model with BitLLM for Edge-Devices
BitTP introduces a lightweight trajectory prediction model for edge devices by converting an LLM-based predictor into a bitlinear architecture. The method employs weight-only quantization to 1.58-bit (BitTP-Weight) while maintaining full-precision activations to avoid degradation in spatio-temporal reasoning. Empirical results show that BitTP-Weight improves prediction quality over the full-precision BF16 LLM baseline, reducing Average Displacement Error (ADE) by 14.29% and Final Displacement Error (FDE) by 20.97%. Additionally, it reduces memory usage and inference latency compared to other quantization methods, demonstrating that careful quantization can act as an effective regularizer for deploying LLM-based reasoning on resource-constrained devices.
trajectory predictionbitlinear architectureweight-only quantizationspatio-temporal reasoningedge devices
Beyond Trajectory Rewards: Step-level Credit Assignment for Agentic Search via Graph Modeling
The paper introduces Graph-Distance Contribution Reward (GDCR), a step-level process reward for Agentic Search that quantifies behavioral contributions by measuring entity distances to the answer node in a training-time Entity-Relation graph. It also proposes Step Advantage Policy Optimization (SAPO), which integrates GDCR-derived step-level advantages with trajectory-level outcome advantages. Experiments on four benchmarks demonstrate the method's effectiveness in improving search performance without costly tree sampling.
agentic searchstep-level rewardentity-relation graphpolicy optimizationcredit assignment
FHRFormer: A Self-Supervised Masked Transformer Framework for Fetal Heart Rate Time-Series Inpainting and Forecasting
FHRFormer introduces a self-supervised masked transformer framework for fetal heart rate (FHR) time-series inpainting and forecasting, addressing signal dropout from wearable monitors. The method employs a transformer-based autoencoder to reconstruct missing FHR data while preserving local temporal and spectral characteristics, outperforming traditional interpolation techniques. Results demonstrate robustness across varying missing-data durations, enabling retrospective analysis for AI-based risk prediction and potential real-time integration into wearable devices.
fetal heart ratemasked transformertime-series inpaintingautoencodersignal dropout
Reliable Reasoning with Large Language Models via Preference-Based Maximum Satisfiability
The paper introduces a hybrid reasoning approach combining LLMs with preference-based Maximum Satisfiability (MaxSAT) to address optimization tasks with multiple constraints. Given natural language problem descriptions, LLMs generate Python code encoding constraints as MaxSAT problems, solved by an exact solver; solutions are verified for feasibility and optimality against canonical encodings. Evaluated on three task families using open/closed-access LLMs, the method achieves up to 80% acceptance rates, outperforming direct-answer, chain-of-thought, and program-of-thought baselines in correctness under verified semantics.
maximum satisfiabilitylarge language modelshybrid reasoningconstraint optimizationcode generation
NICE: A Theory-Grounded Diagnostic Benchmark for Social Intelligence of LLMs
The paper introduces NICE, a theory-grounded diagnostic benchmark for evaluating social intelligence in LLMs, addressing gaps in existing benchmarks through a unified framework derived from psychometric principles. The framework organizes social abilities into 4 categories and 11 dimensions, operationalized via 137 context-specific items in Chinese. Evaluation of 5 frontier LLMs reveals higher aggregate accuracy but consistent weaknesses in Communication, localized to multi-turn communication, nonverbal communication, and synchrony, enabling fine-grained diagnosis of socially consequential model limitations.
social intelligencediagnostic benchmarkpsychometric principlesmulti-turn communicationnonverbal communication
Notation Matters: A Benchmark Study of Token-Optimized Formats in Agentic AI Systems
This study benchmarks token-optimized formats TOON and TRON against JSON in agentic AI systems, evaluating their token efficiency and accuracy in end-to-end workflows. Using four agentic benchmarks (BFCL, MCPToolBenchPP, MCP-Universe, StableToolBench) and five open-weight LLMs, the authors decouple input and output compression to independently measure comprehension and generation. TRON reduces tokens by up to 27% with accuracy within 14pp of JSON, while TOON achieves up to 18% reduction at a 9pp accuracy cost but exhibits cascading parsing failures and collapses parallel tool-call output.
token-optimized formatsagentic ai systemsend-to-end workflowsinput compressionoutput compression
From Prompts to Context: An Ontology-Driven Framework for Human-Generative AI Collaboration
The paper introduces From Prompts to Context, an ontology-driven framework for modeling Human-Generative AI collaboration. The framework leverages the Contextual Collaboration AI Ontology (CCAI) to represent tasks, agent roles, resources, and constraints as a machine-interpretable vocabulary. It integrates populated CCAI instances with SPARQL-based context retrieval to transform ephemeral prompt-response interactions into structured, queryable collaboration traces. A case study in software development demonstrates the framework's ability to enhance task context explicitness, traceability of AI-generated contributions, and transparency in collaborative practices. Results indicate improved accountability and documentation across requirements analysis, design, implementation, and testing phases.
ontology-driven frameworkcontextual collaborationsparql-based retrievalmachine-interpretable vocabularycollaboration traces
EviLink: Multi-Path Schema Linking with Uncertainty-Guided Evidence Acquisition for Large-Scale Text-to-SQL
EviLink introduces uncertainty-guided multi-path schema linking for Text-to-SQL, addressing the challenge of identifying sufficient schema context from ambiguous databases. The method reframes schema linking as uncertainty-aware inference over multiple SQL paths, combining multi-hypothesis schema grounding with evidence acquisition focused on uncertain items. Evaluated on BIRD-Dev and Spider2-Snow, EviLink achieves 90.15% field-level strict recall with 123.30K average tokens, improving downstream SQL generation efficiency.
schema linkingtext-to-sqluncertainty-guidedmulti-hypothesisevidence acquisition
GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents
We introduce GRASP (Gated Regression-Aware Skill Proposer), a method for self-improving LLM agents that treats improvement as a sequence of edits to a bounded skill library, admitting candidates only if they produce a net improvement under a hard regression budget. GRASP evaluates proposed skills on a balanced held-out probe, ensuring no regression in previously correct behavior. Evaluated across five base models (gpt-oss-120b, DeepSeek V4 Flash, Gemini 3.1 Flash Lite, GPT-4.1, GPT-5.4) on FHIR-based clinical benchmarks, GRASP lifts gpt-oss-120b from 40.6% to 88.8% accuracy, outperforming baselines by up to 21.0 points. The method generalizes beyond clinical domains, improving agents in three of four non-clinical environments.
llm agentsskill libraryregression budgetclinical benchmarksself-improvement
Opir: Efficient Multi-Task Safety Classification for Toxicity, Jailbreaks, Hate Speech, and Harmful Content
The paper introduces Opir, a family of efficient multi-task safety classifiers for LLM applications, addressing toxicity, jailbreaks, hate speech, and harmful content detection. Built on GLiClass architecture, Opir includes models for binary classification, multi-label toxicity, jailbreak detection, and zero-shot categorization, with edge variants under 100M parameters. Trained on a 996-category taxonomy using diverse data sources, Opir outperforms eight baselines across 12 safety-classification and 17 category tasks while maintaining a smaller deployment footprint.
multi-task classificationsafety filteringjailbreak detectiongli-class architecturetoxic language
OccamToken: Efficient VLM Inference with Training-Free and Budget-Adaptive Token Pruning
The paper introduces OccamToken, a training-free framework for efficient vision-language model (VLM) inference through adaptive token pruning. The method replaces absolute token ranking with register-anchored relative evidence testing, evaluating whether visual tokens provide information beyond register-based references. This approach enables both image-adaptive redundancy pruning and query-adaptive relevance pruning via dynamic thresholds derived from register attention. Evaluated on LLaVA-NeXT, LLaVA-v1.5, and Qwen3-VL, OccamToken reduces visual tokens from 2,880 to ~40 while preserving 93% accuracy, demonstrating stable compression even at 1.4% retention.
vision-language modelstoken pruningregister-anchored testinginference efficiencyattention sinks
TRACE: Toulmin-based Reasoning Assessment through Constructive Elements for LLM CoT Evaluation
TRACE introduces a novel metric for evaluating Chain-of-Thought (CoT) reasoning processes in large language models (LLMs), addressing the limitations of final-answer accuracy and surface-level statistics. The method integrates Toulmin's argumentation theory with Flavell's metacognitive framework to assess reasoning structure. Experiments on 26,300 QA samples across 7 reasoning models demonstrate a strong correlation (r=0.74) with benchmark accuracy, and TRACE outperforms accuracy-only baselines as a reinforcement learning reward signal. These findings highlight that logically sound reasoning leads to higher-quality answers, positioning TRACE as a complementary metric for open-ended output evaluation.
chain-of-thoughttoulmin's argumentationmetacognitive frameworkreasoning structurereinforcement learning
PTCG-Bench: Can LLM Agents Master Pokémon Trading Card Game?
The authors introduce PTCG-Bench, a benchmark for evaluating LLM agents in the Pokémon Trading Card Game (PTCG) across two dimensions: decision-making in complex environments and self-evolution through accumulated experience. The benchmark includes a modular harness ablation to isolate agent performance from model capability. Experimental results indicate that while LLM agents achieve non-trivial gameplay performance, sustained self-evolution remains challenging, and performance is sensitive to harness design. The work aims to advance research on harness-aware and self-evolving agents in interactive environments.
llm agentsself-evolutionbenchmarkdecision-makingmodular harness
Think Fast, Talk Smart: Partitioning Deterministic and Neural Computation for Structured Health Text Generation
The paper proposes Think Fast, Talk Smart, a hybrid pipeline for structured health text generation that partitions deterministic computation and neural generation. The method uses deterministic code for recurring analysis tasks before a single bounded LLM call, applied to sleep-health insights from 280 user-nights. Evaluations across six models show reduced numeric error (vs. zero-shot/few-shot baselines), improved instruction compliance, and lower operational costs. Layer replacement experiments demonstrate that LLM-based components introduce specific failures: numeric comparison errors, policy selection degradation, and unsupported causal claims. The findings support a design principle of delegating recurring analysis to code while restricting LLMs to bounded factual expression.
structured text generationdeterministic computationhealth informaticsllm promptingerror analysis
LLM-Evolved Domain-Independent Heuristics for Symbolic AI Planning
The paper introduces the first domain-independent heuristics for symbolic AI planning generated via LLM-evolved C++ programs, surpassing hand-engineered baselines. Using MAP-Elites evolutionary search with fitness scores blending coverage and solving time, the method mutates parent heuristics (including blind and FF variants) while tracking informedness-speed tradeoffs. Evolved heuristics achieve superior task-solving performance on unseen domains, forming a Pareto-optimal frontier, with blind heuristic seeding outperforming FF seeding. The C++ implementations integrate seamlessly into existing planners while preserving soundness and completeness guarantees.
symbolic ai planningdomain-independent heuristicsmap-elitesllm-evolved programsinformedness-speed tradeoff
The Sample Complexity of Multiclass and Sparse Contextual Bandits
(No summary returned.)
VikingMem: A Memory Base Management System for Stateful LLM-based Applications
The paper introduces VikingMem, a Memory Base Management System for stateful LLM applications, addressing limitations of existing memory approaches through selective extraction, stateful evolution, and generalizable abstraction. The system implements event-centric memory extraction and dynamic entity updates on the VikingDB vector engine, employing temporal compression and time-weighted recall for progressive summarization. Evaluations show VikingMem outperforms baselines by up to 30% in retrieval effectiveness while maintaining low latency.
memory basestateful evolutiontemporal compressiontime-weighted recallvector engine
Predicting Causal Effects from Natural Language Queries using Structured Representations
The paper introduces Query2Effect, a 72,000-question benchmark for predicting causal effects from natural language queries, simulating real-world information needs through varied query specificity. It proposes a two-step framework that first generates structured query representations, then predicts effect sizes via a supervised encoder. Finetuning reduces absolute error by 27-71% versus prompted LLMs, while the structured approach improves out-of-domain generalization by decoupling semantic interpretation from numerical estimation.
causal effect predictionstructured representationnatural language querieseffect size estimationout-of-domain generalization
Entity-Collision: A Stratified Protocol for Attributing Retrieval Lift in Agent Memory
The article introduces entity-collision, a stratified protocol for attributing retrieval lift in agent memory systems by controlling lexical leakage and tag-mixing. The method constructs a BM25 floor where all distractors share answer entity tokens and stratifies queries by discriminator tag, isolating embedder performance. Results from testing 5 tags × 3 embedders × 5 collision degrees reveal MiniLM-384 outperforms others, while BGE-large shows mixed performance, indicating encoder capacity alone is insufficient. The protocol, tested on LongMemEval and LoCoMo, includes reproducible benchmarks and scripts verified by a public registry.
entity-collisionbm25 floorlexical leakagetag-mixingretrieval lift
Beyond Attack Success Rate: Temporal Logit Observability for LLM Safety Failures
The paper introduces Temporal Logit Observability (TLO), a training-free diagnostic for analyzing LLM safety failures by tracking the compliance-refusal margin during decoding. Unlike Attack Success Rate (ASR), TLO reveals distinct temporal patterns in jailbreak attacks by projecting model-attack conditions onto a calibrated 2D plane. Evaluated across four aligned LLMs and three jailbreak paradigms, TLO differentiates attacks with similar ASR and enables an early-stop rule that reduces successful jailbreaks by >50% without false alarms on benign queries. The method correlates with refusal-direction probes from hidden states, demonstrating logit-based observability of failure dynamics.
temporal logit observabilityjailbreak attackscompliance-refusal marginllm safetyattack success rate
COMET: Concept Space Dissection of the Modality Gap in Audio-Text Multimodal Contrastive Embeddings
The paper introduces COMET, a PLS-SVD framework for analyzing the modality gap in Contrastive Language-Audio Pretraining (CLAP) models. It demonstrates that only a small subset of interpretable axes, capturing shared concepts, significantly contributes to similarity computation, and that the mean component only partially explains the gap. A spectral truncation method is proposed to mitigate the gap without training, enabling zero-shot audio captioning to approach supervised performance while reducing embedding dimensionality. Results show preserved performance on retrieval and captioning tasks.
modality gapcontrastive learningpls-svdzero-shot learningaudio-text embeddings
DLM-SWAI: Steering Diffusion Language Models Before They Unmask
We propose DLM-SWAI, a training-free method for steering diffusion language models (DLMs) toward desired textual properties during inference. Unlike existing approaches designed for autoregressive decoding, DLM-SWAI biases token distributions at each denoising step using pre-computed token-level style scores, enabling control without auxiliary models or retraining. Experiments on style and safety tasks demonstrate effective steering while preserving generation quality, with minimal computational overhead. Ablation studies reveal a controllable trade-off between steering strength and fluency, and analysis links steerability to token-level attribute cues.
diffusion language modelsdenoising steptoken-level style scoressteering strengthinference-time control
Improving Collaborative Storytelling with a Multi-Agent Framework Based on Large Language Models
The paper introduces a multi-agent framework for collaborative storytelling between children and LLMs via physical board games, focusing on ludic co-creation. The method employs an iterative Writer-Editor process, where one LLM (e.g., GPT-3) generates stories and another evaluates and refines them through feedback loops. Simulation results demonstrate that this iterative refinement consistently improves narrative quality, with a small number of steps (exact count unspecified) sufficient for high-quality outputs in interactive settings.
co-creationlarge language modelsmulti-agent frameworkiterative refinementinteractive storytelling
Learning Context-Conditioned Predicate Semantics via Prototype Feedback
AlignG introduces a novel approach for learning context-conditioned predicate semantics in scene graph generation, addressing the challenge of polysemous predicates whose meanings vary across contexts. The method infers context-conditioned predicate semantics from relation candidates within each image and uses prototype feedback to recalibrate relation representations, anchored to global semantic centers to prevent semantic drift. Experiments on VG-150 and GQA-200 demonstrate consistent improvements, with F@100 gains of +1.4 and +2.7 respectively under SGDet. Visualizations reveal coherent context-dependent reorganization of prototypes based on scene evidence.
scene graph generationpolysemous predicatesprototype feedbacksemantic driftcontext-conditioned semantics
HiKEY: Hierarchical Multimodal Retrieval for Open-Domain Document Question Answering
The paper introduces HiKEY, a hierarchical multimodal retrieval framework for Open-domain Document Question Answering (ODQA) that addresses routing failure and evidence fragmentation. HiKEY employs Document Hierarchical Parsing (DHP) to construct a logical heterogeneous graph, enabling global routing and fine-grained retrieval via hierarchical indexing and multimodal fusion. Experimental results show HiKEY improves retrieval recall by up to 12.9% and end-to-end QA performance by up to 6.8% over baselines.
retrieval-augmented generationhierarchical indexingmultimodal fusiondocument hierarchical parsingevidence subgraph
Training Deliberative Monitors for Black-Box Scheming Detection
The paper introduces action-only deliberative monitors, smaller open-weight models trained to detect scheming behavior in autonomous agents without accessing internal reasoning. The method uses a scheming specification to generate structured rationales via a frontier teacher, filters them with a judge, and distills high-quality rationales into monitors using supervised fine-tuning and reinforcement learning. Evaluated on six out-of-distribution benchmarks, monitors trained on Qwen3.5-27B outperform low-cost frontier models (Gemini 3.1 Flash-Lite, GPT-5.4 Nano, Claude Haiku 4.5) and Gemini 2.5 Pro, achieving lower marginal inference cost ($16$--$34\times$ cheaper) than stronger prompted monitors (Gemini 3.1 Pro, GPT-5.4, Claude Sonnet 4.6, Claude Opus 4.6).
deliberative monitorsscheming detectionopen-weight modelssupervised fine-tuningreinforcement learning
Mind-Omni: A Unified Multi-Task Framework for Brain-Vision-Language Modeling via Discrete Diffusion
Mind-Omni introduces a unified multi-task framework for brain-vision-language modeling via discrete diffusion, addressing limitations of single-task approaches. The method employs a Brain Tokenizer to convert continuous brain signals into discrete tokens, enabling cross-modal interactions in a shared semantic space, and incorporates a Brain Question Answering dataset for instruction tuning. Results show state-of-the-art performance across seven tasks, with competitive or superior results to larger specialized models, demonstrating multi-task synergy and advancing neural activity foundation models.
brain-computer interfacesdiscrete diffusionmulti-task learningneural representationsinstruction tuning
Brain-IT-VQA: From Brain Signals to Answers
The paper introduces Brain-IT-VQA, a framework for visual question answering (VQA) from fMRI signals, and NSD-VQA, a new benchmark dataset. Building on the Brain Interaction Transformer (Brain-IT), the method decodes language tokens from brain activity and integrates them with a language model to answer visual questions. The model outperforms previous fMRI-based captioning and VQA approaches, while NSD-VQA provides 20 question-answer pairs per image across 20 controlled categories, enabling reliable evaluation. The benchmark quantifies decodable visual and semantic information from fMRI and analyzes regional brain contributions across question types.
fmrivisual question answeringbrain decodingtransformerneural representations
FinVerBench: Benchmark Validity and Calibration in Large Language Model Financial Statement Verification
The paper introduces FinVerBench, a benchmark for evaluating large language models (LLMs) on financial statement verification tasks, assessing numerical consistency in corporate filings. The benchmark comprises 43 S&P 500 company SEC 10-K XBRL filings, with a four-category error taxonomy (arithmetic, cross-statement linkage, year-over-year, magnitude perturbations). Testing 14 LLM runs (excluding one incomplete Gemini 2.5 Pro run) reveals high false positive rates (95-100% for 9 models) on clean statements, with one model achieving 0% false positives. Rendering choices significantly impact recall, with a calibrated model showing 79.0% recall on rounded data versus 100% on unrounded data, highlighting the task's complexity beyond mere arithmetic.
financial statement verificationlarge language modelserror taxonomynumerical consistencybenchmark validity
GPS-Enhanced Tourist Mobility Modeling with Seasonal Spatial Priors and LLM-Based Activity Chain Generation
The paper proposes a four-stage framework for tourist mobility modeling that integrates seasonal spatial priors and LLM-based activity generation. The method combines: (1) month-conditioned spatial priors from aggregated GPS/survey data, (2) trip extent prediction from demographics, (3) distance-constrained ward sequence assignment, and (4) LLM-generated activity chains under household/spatial constraints. Evaluated on Tokyo tourism data, the framework produces synthetic schedules with ward-level visitation shares aligning within 5% of survey distributions while preserving privacy through aggregated GPS processing.
tourist mobility modelingspatial priorsactivity chain generationllm-based simulationprivacy-preserving aggregation
DeepTool: Scaling Interleaved Deliberation in Tool-Integrated Reasoning via Process-Supervised Reinforcement Learning
DeepTool introduces a novel framework for Tool-Integrated Reasoning (TIR) that enhances LLM capabilities through interleaved deliberation. The method combines a synthesis pipeline for robust trajectory generation with Process-Supervised Reinforcement Learning (GRPO-based) using Action-Centric Process Rewards to supervise intermediate reasoning steps. Experiments show significant performance improvements, elevating Qwen2.5-7B's accuracy on benchmarks like AIME24 (3.2% to 40.4%) and HMMT25 (0.0% to 28.6%), while maintaining token efficiency.
tool-integrated reasoningprocess-supervised rlinterleaved deliberationaction-centric rewardtoken efficiency
Planning with the Views via Scene Self-Exploration
The paper introduces view planning, a novel capability for vision-language models (VLMs) to predict camera movements and compose multi-turn plans in 3D environments. The authors propose ViewSuite, a 3D point-cloud benchmark based on ScanNet, revealing that current VLMs (tested across 13 models) understand single-view transformations but fail at multi-step composition. To address this, they present an iterative framework combining self-exploration and view graph distillation, which transforms exploration trajectories into diverse supervised tasks. This approach elevates Qwen2.5-VL-7B's performance from 2.5% to 47.8%, outperforming GPT-5.4 Pro (18.5%) and Gemini 3.1 Pro (21.4%).
view planningvision-language models3d point-cloudself-explorationview graph distillation
VLA-Pro: Cross-Task Procedural Memory Transfer for Vision-Language-Action Models
VLA-Pro introduces a plug-and-play framework for enhancing cross-task generalization in Vision-Language-Action (VLA) models through procedural memory transfer. The method stores task-specific LoRA adapters as parameterized procedural memories during training and dynamically retrieves/fuses them during inference based on multi-modal context. Experiments on RoboTwin, RLBench, and real-world tasks demonstrate up to 207% relative improvement in simulation and a real-world success rate increase from 5.8% to 65.0%, validating effective experience transfer while maintaining modularity.
vision-language-action modelsprocedural memorylora adapterscross-task generalizationmulti-modal fusion
ParaTool: Shifting Tool Representations from Context to Parameters
ParaTool introduces a framework for parameterizing tool representations in LLMs, eliminating reliance on in-context documentation while maintaining tool-calling capabilities. The method involves three stages: parametric tool pre-training to encapsulate tool knowledge, soft tool selection via a gating network, and joint parametric tool fine-tuning. Evaluations on Stable ToolBench and BFCL show ParaTool outperforms in-context learning baselines in both performance and computational efficiency.
tool callingparameter modulesgating networkin-context learningcomputational complexity
Battery-Sim-Agent: Leveraging LLM-Agent for Inverse Battery Parameter Estimation
The paper introduces Battery-Sim-Agent, a novel framework leveraging a Large Language Model (LLM) agent for inverse battery parameter estimation. The agent operates in a closed loop with a high-fidelity battery simulator, interpreting multi-modal feedback, forming physically-grounded hypotheses, and proposing structured parameter updates, mimicking a human scientist's workflow. Evaluated on a benchmark suite spanning diverse battery chemistries, operating conditions, and difficulty levels, the agent outperforms Bayesian optimization and other black-box optimization baselines in parameter accuracy. The framework demonstrates effectiveness in complex long-horizon degradation fitting tasks and validates practical applicability on real-world battery datasets.
inverse battery parameter estimationlarge language modelblack-box optimizationhigh-fidelity simulatordegradation fitting
Opt-Verifier: Unleashing the Power of LLMs for Optimization Modeling via Dual-Side Verification
Opt-Verifier introduces a dual-side verification framework leveraging large language models (LLMs) to enhance optimization modeling accuracy in operations research. The method employs structure-side verification to ensure alignment between generated models and problem descriptions, and solution-side verification to validate solution correctness and mathematical soundness. Experiments on standard benchmarks demonstrate a 20% improvement in modeling accuracy compared to existing approaches.
optimization modelinglarge language modelsdual-side verificationoperations researchmathematical soundness
Singularity-aware Optimization via Randomized Geometric Probing: Towards Stable Non-smooth Optimization
Singularity-aware Adam (S-Adam) is introduced to stabilize deep learning optimization in non-smooth regimes, addressing gradient chattering caused by conflicting signals in the Clarke subdifferential. S-Adam dynamically modulates step sizes using the Local Geometric Instability (LGI) metric, derived from the variance of randomized directional derivatives, and incorporates an adaptive damping mechanism exp(-λρ) to decelerate updates in high-instability regions. Rigorous convergence analysis proves S-Adam converges almost surely to (δ,ε)-Clarke stationary points at the optimal O(1/√T) rate. Empirical evaluations on Quantization-Aware Training (QAT) and high-noise small-batch learning show S-Adam outperforms AdamW and Prox-SGD, achieving accuracy gains of up to 6% on CIFAR-100 and 3% on TinyImageNet.
clarke subdifferentiallocal geometric instabilityadaptive dampingquantization-aware traininggradient chattering
SCOPE: A Lightweight-training LLM Framework for Air Traffic Control Readback Monitoring
The paper introduces SCOPE, a lightweight-training LLM framework for Air Traffic Control (ATC) readback monitoring, addressing deployment and computational barriers of existing approaches. SCOPE combines a plug-in open-set classifier with in-context learning on a frozen LLM, enhancing both efficiency and accuracy. Evaluated on a semi-synthetic dataset, SCOPE achieves 91.05% accuracy in open-set detection and corrects 96.63% of anomalous readbacks under few-shot settings, outperforming baselines while providing interpretable decisions.
large language modelsin-context learningopen-set classificationair traffic controlreadback monitoring
GiPL: Generative augmented iterative Pseudo-Labeling for Cross-Domain Few-Shot Object Detection
The paper introduces GiPL, a two-branch framework for Cross-Domain Few-Shot Object Detection (CD-FSOD) addressing insufficient support set utilization and overfitting. The first branch employs iterative pseudo-label self-training to generate reliable annotations from zero-shot inference, fused with ground-truth labels. The second branch uses vision-language models for generative data augmentation, synthesizing domain-aligned multi-object images. Evaluations on RUOD, CARPK, and CarDD datasets under 1/5/10-shot settings show GiPL outperforms state-of-the-art methods with significant gains.
cross-domainfew-shotobject detectionpseudo-labelinggenerative augmentation
UI-KOBE: Knowledge-Oriented Behavior Exploration for Lightweight Graph-Guided GUI Agents
UI-KOBE introduces a knowledge-oriented behavior exploration framework to enhance lightweight mobile GUI agents by leveraging reusable app-specific graph knowledge. The method autonomously constructs an app knowledge graph with UI states as nodes and transitions as edges, then guides a lightweight agent during runtime to select actions based on the current state and task. This approach reduces end-to-end planning demands, enabling more effective task execution while maintaining efficiency, interpretability, and privacy for on-device deployment.
mobile gui agentsknowledge graphlightweight modelson-device deploymentbehavior exploration
GUITestScape: Towards Open-set Evaluation on Exploratory GUI Testing
GUITestScape introduces an open-set evaluation framework for exploratory GUI testing, addressing limitations in current benchmarks that overlook display defects and rely on predefined defect annotations. The framework comprises 61 Android applications with 508 preset defects spanning interaction and display types, alongside GUIJudge, an evaluator that decomposes testing trajectories into independently diagnosable capabilities. Experiments show GUIJudge enables reliable process-aware evaluation, outperforming baselines, and reveals detection as the critical bottleneck for existing models. Integration of GUIJudge's verifiers enhances detection performance without retraining, demonstrating its utility in improving GUI testing agents.
exploratory gui testingopen-set evaluationandroid applicationsprocess-aware evaluationdetection bottleneck
Temporal Motif-aware Graph Test-time Adaptation for OOD Blockchain Anomaly Detection
The paper introduces TEMG-TTA, a novel framework for out-of-distribution blockchain anomaly detection that addresses adversarial pattern evolution and transaction semantic variability. The method captures 3-node temporal motif distributions using an efficient computational mechanism and implements a test-time adaptation strategy to share common patterns between training and testing graphs. Evaluated on 5 real-world datasets, TEMG-TTA outperforms state-of-the-art graph anomaly detection approaches by an average of 54.88%. Case studies demonstrate its ability to characterize complex transaction patterns of anomalous addresses, validating the technical design.
temporal motifgraph anomaly detectiontest-time adaptationout-of-distributionblockchain
KBF: Knowledge Boundary as Fingerprint for Language Model and Black-Box API Auditing
The paper introduces KBF, a black-box auditing protocol for verifying language model APIs by fingerprinting their knowledge boundaries. The method detects model substitutions through stable numerical recall patterns near the boundary of a model's knowledge. Evaluations on 16 production LLM endpoints show 100% detection of economically relevant substitutions (155/155) with zero false positives, robustness to deployment variations, and sensitivity to mixed-routing attacks (5-10% substitution). A shadow API audit revealed 7/27 platform model cells exhibited statistical inconsistencies, particularly affecting premium Claude endpoints.
knowledge boundaryblack-box auditingmodel fingerprintingnumerical recallmixed-routing attacks
DeepSurvey: Enhancing Analytical Depth and Citation Reliability in Automated Survey Generation
DeepSurvey introduces an agentic system for automated survey generation that enhances analytical depth and citation reliability. It extracts structured keynotes from full-text papers, models cross-paper relationships via clustering and comparative analysis, and integrates code-repository analysis for implementation-level details. Citation reliability is fortified through citation-graph expansion, hybrid filtering, evidence-constrained citation assignment, and multi-granularity agentic refinement. Experiments demonstrate DeepSurvey achieves the highest content score (8.644/10), improves citation recall and precision by 12.3% and 9.3% over baselines, generalizes robustly across domains (0.14 vs 0.22 to 0.69 CS-to-non-CS drop), and is preferred by domain experts (83.3% overall quality, 100% content depth).
agentic systemcitation-graph expansioncross-paper relationshipsmulti-granularity refinementcode-repository analysis
Network Optimization Aspects of Autonomous Vehicles: Challenges and Future Directions
The article contributes a comprehensive review of network optimization challenges in Connected and Autonomous Vehicles (CAVs), addressing public misconceptions and outlining future directions. It employs multidisciplinary methods, including cooperative perception, to analyze CAV network optimization. Drawing on extensive experience, the authors present insights, relevant use-cases, and experimental results to advance understanding in this domain.
connected and autonomous vehiclesnetwork optimizationcooperative perceptionmultidisciplinary methodsuse-cases
MINDGAMES: A Live Arena for Evaluating Social and Strategic Reasoning in Multi-Agent LLMs
The paper introduces MINDGAMES, a multi-game evaluation platform for assessing social and strategic reasoning in LLM agents through sustained multi-agent interactions. The platform operationalizes theory-of-mind demands across four games (Colonel Blotto, Iterated Prisoner's Dilemma, Codenames, Secret Mafia) with unified interfaces, TrueSkill ratings, and trajectory logging. Analysis of 944 agents from 76 teams reveals limitations in rule adherence, structural scaffolding dependence, and leaderboard validity variations, with error-survival confounds observed in Secret Mafia. The authors release a dataset of 29,571 games and MG-Ref, a deterministic offline tournament protocol.
multi-agent llmstheory of mindstrategic reasoningtrueskill ratingerror-survival confound
Xetrieval: Mechanistically Explaining Dense Retrieval
Xetrieval introduces a mechanistic framework for explaining dense retrieval by operating directly on embedding representations. The method employs a lightweight reasoning internalizer to enrich sentence embeddings with Chain-of-Thought-like reasoning in a single forward pass, avoiding costly autoregressive generation. It then decomposes these embeddings into sparse, interpretable features with natural language descriptions, enabling feature-level explanations of retrieval decisions through aggregation across document views. Experiments demonstrate that Xetrieval uncovers coherent features, improves pair-level intervention effects, and supports task-level feature steering across diverse retrievers and benchmarks.
dense retrievalembedding-level explanationreasoning internalizersparse featureschain-of-thought
Source-Grounded Semantic Reinforcement Learning for Low-Resource Target-Language Generation
The paper introduces Source-Grounded Semantic Reinforcement Learning (SG-SRL), a framework for low-resource target-language generation that leverages high-resource source-language monolingual data via cross-lingual semantic supervision. SG-SRL employs reference-free reinforcement learning with a cross-lingual semantic reward model (instantiated as a reranker) to score semantic relevance between source input and target output, followed by a lightweight recovery stage using minimal parallel data to address verbosity-based reward hacking. Experiments on Chinese-to-Thai generation demonstrate improved semantic grounding and factual coverage over cold-start supervised fine-tuning, with additional analyses validating generalization to long-form transfer and the use of encoder-based rewards in low-resource settings.
low-resource generationcross-lingual semantic rewardreference-free rlverbosity-based reward hackingencoder-based reranker
Quotient DAGs for Off-Policy Evaluation:Forward-Flow Importance Sampling and Exact Slate Propensities
The paper introduces Forward-DP, a dynamic programming method for exact unordered slate propensity computation in off-policy evaluation of autoregressive slate recommenders. By leveraging a quotient-DAG framework that merges equivalent histories and employs target-to-behavior forward-flow ratios, the method avoids factorial enumeration of generation orders. This enables efficient computation of exact propensities for context-dependent autoregressive slate loggers. The approach reduces nuisance variance and bridges the computational gap in standard importance sampling, facilitating practical propensity-based evaluation and model selection.
off-policy evaluationquotient-dagforward-flow ratiosslate recommendationdynamic programming
The New Pro Se: Generative AI and the Surge in Federal Civil Self-Representation
This study examines the impact of generative AI on pro se litigation in federal civil cases, analyzing 2.8 million filings from FY2008-2025. The pro se plaintiff rate increased from 11.33% pre-GenAI to 16.94% post-GenAI, a 5.61 percentage-point rise. Using stylometric AI detection, 13.9% of post-GenAI complaints showed AI-consistent drafting, characterized by higher citation density and association with first-time filers. AI-flagged complaints were more likely to be dismissed early, with no improvement in win rates. Findings highlight disparities in access to justice and court efficiency.
generative aipro se litigationstylometric analysiscivil rightslegal efficacy
The Curse of Helpfulness: Inverse Scaling Law in Robustness to Distractor Instructions via DistractionIF
The paper introduces DistractionIF, a benchmark for evaluating LLM robustness against instruction-like semantic noise in reference text, revealing an inverse scaling law where larger models exhibit up to 30-point performance drops due to over-interpreting distractors as instructions. Mechanistic analysis via perplexity shows scaling erodes the probabilistic boundary between robust and distracted behaviors. The authors demonstrate that Group Relative Policy Optimization (GRPO) improves robustness by 15.5% without compromising general instruction-following capability, establishing reinforcement learning as a viable solution for enforcing data-instruction separation.
inverse scaling lawretrieval-augmented generationdistractor instructionsgroup relative policy optimizationinstruction-following robustness
AnyMo: Scaling Any-Modality Conditional Motion Generation with Masked Modeling
We introduce AnyMo, a unified multimodal framework for conditional human motion generation, addressing limitations in cross-modal interactions and scalability. The method combines a Residual FSQ-based motion tokenizer with a scalable masked modeling transformer, trained on OmniHuMo, a large-scale dataset comprising 5,000 hours of motion and 3.2 million sequences with aligned multimodal annotations (text, speech, music, trajectory). Experiments demonstrate AnyMo's capability for high-fidelity synthesis with flexible control over spatial and stylistic attributes under arbitrary modality combinations.
residual fsqmasked modeling transformermultimodal annotationsmotion tokenizerconditional motion generation
PhoneWorld: Scaling Phone-Use Agent Environments
PhoneWorld introduces a scalable pipeline for constructing phone-use environments by converting real GUI trajectories and screenshots into executable tasks, automatic verifiers, and training rollouts. The method leverages real trajectories to identify relevant screens, their connections, state-changing interactions, and verifiable user goals, enabling the creation of mock Android apps with mutable state. PhoneWorld currently covers 34 apps across 16 domains, including search, shopping, and social interaction. Empirical results show that integrating PhoneWorld supervision improves performance across four benchmarks: HYMobileBench (+17.7), AndroidControl (+6.0), AndroidWorld (+14.7), and PhoneWorld (+52.5). Scaling experiments demonstrate that increasing supervision and app coverage further enhances performance.
gui trajectoriesmock android appsautomatic verifierstraining rolloutsscalable pipeline
VitalAgent: A Tool-Augmented Agent for Reactive and Proactive Physiological Monitoring over Wearable Health Data
VitalAgent introduces a tool-augmented agentic framework for ECG/PPG-based mHealth, enabling both reactive question answering and proactive monitoring over long-term physiological signals. The framework leverages a longitudinal physiological memory and a tool-augmented reasoning interface for dynamic computation over raw signals. Evaluated on VitalBench, a benchmark dataset with 1,862 QA pairs and 90.2 hours of ECG/PPG recordings, VitalAgent achieves over 30% improvement in reactive tasks compared to prompt-based and ReAct baselines, demonstrating effective proactive alert monitoring and dynamic tool use.
tool-augmentedphysiological monitoringecg/ppglongitudinal memoryreactive qa
📰 Industry Media (8)
How the Pope’s Magnifica Humanitas offers a template for individuals to meet the AI moment
Pope Leo XIV's encyclical 'Magnifica Humanitas' frames AI governance as a moral imperative, arguing that technology is never neutral and requires collective human responsibility. The document contrasts unchecked technological expansion (symbolized by the Tower of Babel) with collaborative rebuilding (inspired by Nehemiah), emphasizing shareholder activism as a regulatory alternative amid inadequate institutional oversight. Faith-based and secular investors managing $400B+ assets have filed proxy resolutions demanding AI transparency, risk assessments, and ethical deployment at major tech firms, addressing concerns ranging from military targeting to healthcare and environmental impacts.
ai governanceproxy resolutionsethical deploymentshareholder activisminstitutional oversight
Meet mKernel: A Multi-GPU, Multi-Node Fused Kernel Library for GPU-Driven Communication
mKernel introduces a library of persistent CUDA kernels that fuse intra-node NVLink communication, inter-node RDMA, and dense compute into a single kernel, addressing GPU communication overhead in multi-GPU, multi-node setups. The library enables fine-grained intra-kernel overlap at tile/chunk granularity and leverages GPU-driven networking via libibverbs, eliminating dependencies on NCCL or NVSHMEM. Five fused kernels—AllGather+GEMM, GEMM+AllReduce, MoE Dispatch+GEMM, Ring Attention, and GEMM+ReduceScatter—were evaluated on 2-node × 8-H200 clusters with AWS EFA and ConnectX-7 backends, demonstrating potential to reduce communication bottlenecks in production AI workloads.
persistent cuda kernelsintra-node nvlinkinter-node rdmafine-grained overlapgpu-driven networking
Hexo Labs Open-Sources SIA: A Self-Improving Agent That Updates Both the Harness and the Model Weights
Hexo Labs introduces SIA, a self-improving AI framework that jointly optimizes both the agent's scaffold (system prompt, tool logic) and model weights via LoRA adapters (rank 32). The system employs three LLM components: a Meta-Agent for scaffold generation, a Task-Specific Agent for execution, and a Feedback-Agent that dynamically selects between scaffold updates or weight tuning using task-specific RL algorithms (PPO, GRPO, etc.). Evaluated on LawBench (70.1% accuracy), AlphaEvolve TriMul (14.02× speedup), and RNA denoising (0.289 MSE), SIA-W+H outperformed scaffold-only (SIA-H) and prior SOTA across domains, though co-optimization risks Goodhart effects.
self-improving agentlora adaptersscaffold optimizationentropic advantage weightinggoodhart effects
How to Design an End-to-End Ansible Automation Lab with Playbooks, Inventories, Roles, Vault, Dynamic Inventory, and Custom Modules
The tutorial presents a comprehensive Ansible automation lab design, demonstrating end-to-end configuration management using local execution. Methodologically, it implements ansible-core with static/dynamic inventories, custom modules (system_report), Jinja2 templates, and Ansible Vault for secret management. Results include idempotent playbook execution (verified through --check), role-based web server deployment, and dynamic inventory integration, achieving a production-like environment without remote infrastructure.
ansible-corejinja2 templatesdynamic inventoryidempotent executionansible vault
Liquid AI Releases LFM2.5-8B-A1B: An On-Device MoE Model With 8.3B Total and 1.5B Active Parameters
Liquid AI introduces LFM2.5-8B-A1B, an on-device Mixture-of-Experts (MoE) model with 8.3B total and 1.5B active parameters, optimized for tool calling and multilingual reasoning. The architecture combines sparse MoE, GQA, and gated short convolution blocks, featuring a 128K context window and improved tokenization for nine languages. Training involved extended tokenizer adaptation, staged context growth, and targeted RL to reduce hallucinations and reasoning loops. Benchmarks show significant gains over its predecessor, including a +56.01 improvement in AA-Omniscience Non-Hallucination Rate and 18.5K tokens/s throughput on H100 GPUs.
mixture-of-expertson-device inferencesparse activationcontext windowreinforcement learning
Anthropic Ships Claude Opus 4.8 Alongside Dynamic Workflows and Cheaper Fast Mode, With Workflows Capped at 1,000 Subagents
Anthropic released Claude Opus 4.8 with two key updates: dynamic workflows and cheaper fast mode. Dynamic workflows enable parallel subagent orchestration via JavaScript scripts, decoupling task planning from Claude's context window (16 concurrent agents, 1,000 total cap). The Bun rewrite case study demonstrated 99.8% test suite preservation during a Zig-to-Rust migration. Fast mode offers 2.5x faster token generation at reduced cost ($30/MTok for Opus 4.8). Both features operate as research previews with increased token consumption.
dynamic workflowssubagent orchestrationopus 4.8fast modetoken generation
Scaling safe enterprise AI with OpenAI governance frameworks
OpenAI introduces the Frontier Governance Framework (FGF), a structured approach for enterprise-scale AI deployment aligned with EU and California regulations. The framework categorizes systemic risks (e.g., cyber offense, CBRN threats) into tiers, with Tier 3 representing autonomous model capabilities exceeding human expertise. It prescribes ISO/SOC-compliant security measures, sandboxed execution, and mandatory six-month safety reviews for high-capacity models. Enterprises can operationalize these guidelines through deterministic fail-safes, encrypted middleware, and AI Safety Incident Response Plans (AIRP) mirroring OpenAI's protocols.
frontier governance frameworksystemic risk tiersiso 27001 complianceretrieval-augmented generationsafety incident response
Anthropic releases Claude Opus 4.8
Anthropic released Claude Opus 4.8, an upgraded LLM featuring improved coding, agentic workflows, and reasoning capabilities. Key innovations include dynamic workflows for large codebases (handling 100k+ LOC), live message array updates in the Messages API for mid-task instruction modification, and configurable effort controls (default/xhigh) to manage token consumption-performance tradeoffs. Benchmark tests show 4x reduction in undetected flawed code output versus Opus 4.7, with comparable safety metrics to Claude Mythos Preview. The model achieves 2.5x speed in fast mode ($10/$50 per million I/O tokens) while maintaining coding performance parity with GPT-5.5 in internal evaluations. Enterprise features include expanded rate limits and research-preview toolchain integrations.
agentic workflowsdynamic workflowsmessages apieffort controltoken-based billing
Generated automatically at 2026-05-29 21:27 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.