Daily Digest — 2026-09-14
6 items · 6 industry media
🏛️ Research Labs
No new items today.
📜 arXiv Papers
No new items today.
📰 Industry Media (6)
Hierarchical NeRF with JAX3D for Volumetric Rendering, Novel-View Synthesis, and 3D Reconstruction
The tutorial presents a hierarchical Neural Radiance Field (NeRF) implementation using JAX3D for volumetric rendering and novel-view synthesis. The method constructs a synthetic multi-view dataset with volumetric geometry and view-dependent radiance, then trains a NeRF model with positional encoding, skip connections, and separate coarse/fine networks. Results demonstrate 360-degree rendering, depth visualization, and geometry extraction via marching cubes, evaluated using PSNR and sampling diagnostics.
neural radiance fieldsvolumetric renderingnovel-view synthesisjax3dpositional encoding
A Princeton Researcher Proposes Recurrent Looped Transformer (RLT) that Carries Decoder State across Every Token, Fixing 96 Blocks per Token with Unbounded Temporal Depth
The Recurrent Looped Transformer (RLT) proposes a decoder-only architecture where the final hidden state and sliding-window attention (SWA) cache propagate across all tokens, enabling unbounded temporal depth with fixed per-token compute. The design combines a causal encoder (parallel token processing) with a recurrent decoder (sequential state updates), using 48 tied encoder-decoder layers (96 logical blocks/token). Key innovations include layerwise SWA cache retention, encoder-decoder weight sharing, and RL-compatible state transitions requiring full recomputation for replay. No empirical results are reported; hardware co-design targets batched kernel fusion but avoids parallel scan assumptions.
recurrent looped transformersliding-window attentionunbounded temporal depthdecoder state propagationrl replay
AWS Introduces Pizza Bot: An Open Source Inbox for Background AI Agents
AWS introduced Pizza Bot, an open-source, self-hosted inbox for background AI agents that decouples task initiation from result review. The system uses DeepAgents and LangGraph for stateful execution, with LangGraph checkpoints preserving thread state and approval pauses. It supports multiple model providers (Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter, Ollama) and offers scoped skills via SKILL.md files with configurable approval controls (interruptOn, allowedDecisions). Tasks can be triggered manually, via cron, or webhooks, and continue after client disconnection if a backend remains running. Internally deployed to over 2,000 users, Pizza Bot organizes results into an email-style inbox with All, Unread, and Action views. The code is Apache 2.0 licensed with desktop builds for macOS, Windows, and Linux.
deepagentslanggraphstateful executionapproval controlsasynchronous tasks
Context Engineering Inside the Harness: 4 Mechanisms That Beat Context Overflow and Goal Loss on Long-Horizon Tasks
The article identifies four harness-level mechanisms to mitigate context overflow and goal loss in long-horizon LLM agent tasks: (1) context budgeting and offloading (e.g., truncating tool calls at 85% window capacity), (2) compaction via summarization with preserved fields (e.g., Claude Code retains architectural decisions), (3) todo-state recitation to bias attention (e.g., Manus's dynamic todo.md), and (4) persistent memory strategies (e.g., AgentCore's background extraction). Evaluations show tradeoffs: LangChain's structured summaries improve goal retention, while ETH Zurich found persistent files increase inference costs by 19-23%. Subagent delegation (AWS AgentCore) reduces runtime by 3x versus sequential processing.
context budgetingcompactiontodo-state recitationsubagent delegationattention budget
Implementation of Machine Learning Workflows with NVIDIA cuML, RAPIDS, GPU Benchmarking, Explainability, Clustering, and Model Inference
The tutorial demonstrates GPU-accelerated machine learning workflows using NVIDIA cuML and RAPIDS, comparing performance against CPU-based scikit-learn implementations. Methods include configuring GPU environments, benchmarking PCA/K-Means/DBSCAN algorithms with synchronized timing, and validating SHAP explanations. Results show significant speedups (e.g., 10-50× for nearest-neighbor search and random forests) while maintaining numerical parity, with zero-copy interoperability between CuPy, cuDF, and cuML components.
gpu-accelerationcumlrapidsbenchmarkingshap
Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost
Cognition introduces SWE-2, a reinforcement learning (RL)-post-trained coding model derived from Moonshot AI's 2.8T-parameter Kimi K3, achieving 50.0% on FrontierCode 1.1 Main (within 1 point of Fable 5.1) at 64% lower cost. The model features selectable reasoning-effort levels trained jointly via Pareto-informed cost penalties, optimizing both success rate (S) and cost (C) with a linear reward R = S − λC. SWE-2 reduces median steps per run by 58% versus SWE-1.7 and outperforms its base model by 5–6 points on benchmarks like DeepSWE 1.1 (73.0%) and Terminal-Bench 2.1 (92.8%), though it lags on Terminal-Bench 4 (27.3%). Deployed exclusively within Cognition's Devin ecosystem, SWE-2 employs DSpark speculative decoding and NVFP4/FP8 kernels for efficiency.
reinforcement learningpost-trainingspeculative decodingparameter-efficientbenchmarking
Generated automatically at 2026-09-13 21:16 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.