Daily Digest — 2026-09-08
9 items · 1 research labs, 8 industry media
🏛️ Research Labs (1)
Supporting independent journalism in Ukraine
OpenAI, WAN-IFRA, and AIRPPU launched a joint initiative to bolster Ukrainian independent journalism through AI adoption, targeting editorial workflows, audience engagement, and revenue generation. The program comprises two components: the Newsroom AI Masterclass Series, offering expert-led practical training, and the Newsroom AI Catalyst, providing hands-on support to 10 news organizations for piloting tailored AI solutions. Participants receive OpenAI API credits, with the Masterclass starting August 5, 2026, and the Catalyst launching September 17, 2026, aiming to enhance sustainability and resilience during conflict.
ai adoptioneditorial workflowsaudience engagementapi creditsnewsroom resilience
📜 arXiv Papers
No new items today.
📰 Industry Media (8)
OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device
OpenBMB introduces MiniCPM5-2B, a 2.52B-parameter dense causal LM with 131K context length, achieving 53.9 average score across 34 benchmarks. The model employs grouped-query attention (16 query, 2 KV heads) and follows a LlamaForCausalLM architecture for broad compatibility. Training involves UltraData-tiered pre-training, 400B-token SFT, multi-expert RL (JustRL II), and on-policy distillation (OPD) with reverse KL divergence. It outperforms Qwen3.5-4B (51.1) and excels in tool use (97.1 on τ²-Bench Telecom) and code (69.1 LiveCodeBench), though trails on MMLU-Pro (70.8 vs 78.0). OpenBMB releases weights (Apache 2.0), datasets (Ultra-FineWeb, UltraX), and intermediate checkpoints.
grouped-query attentionon-policy distillationultradatajustrl iireverse kl divergence
Axis Robotics Releases AXIS: A Browser-Based Data Engine With 207 Robot Manipulation Tasks and 50,129 Trajectories
Axis Robotics introduces AXIS, a browser-based data engine for scalable robot manipulation dataset collection, featuring 207 tasks and 50,129 trajectories. The system employs a MuJoCo WebAssembly frontend for teleoperation, with physics stepping and rendering handled by 8x RTX 4090 GPUs, while training/evaluation uses 8x A100 GPUs. TaskGen generates tasks via language instruction decomposition and 3D mesh synthesis, validated by a layout supervisor. Continual pretraining on AXIS improves a PaliGemma Gemma-2B model (π0.5) from 83.9 to 88.8 on LIBERO-Plus, with a 63.9% reduction in mean acceleration and 80.8% in jerk, albeit with 86.2% replay success. Sensor Noise and Camera perturbations show the largest gains (+13.7 and +11.3).
mujoco wasmtaskgenpalgemmalibero-plussavitzky-golay
IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B
The Institute of Foundation Models (IFM) released K2 Horizon, a suite of six open-source models (0.9B to 375B parameters) under Apache 2.0, featuring shared architecture and tooling. Key innovations include Mixture-of-Value Attention (MoVA), which extends MoE routing to attention heads (36B total, ~4B active/token), and Uno, a LoRA adapter enabling 3× lossless decoding speedup via diffusion distillation. Models were pretrained on 20T tokens (17% reasoning trajectories, 10T synthetic), with post-training data integrated mid-training. The 375B variant scored 70.2 on Terminal-Bench 2.1 (66.9 after reward-hacking audit), while smaller models (0.9B-7B) achieved SOTA on SWE-bench and HumanEval+. IFM disclosed full training logs, code, and adversarial benchmarks.
mixture-of-value attentiondiffusion distillationreward-hacking auditsparse attentionin-context learning
H Company Releases NeoMME: A Family of 260M and 800M Single-Tower Multimodal Encoders That Drop the Vision Tower and Causal Decoder
H Company introduces NeoMME, a family of single-tower multimodal encoders (260M and 800M parameters) that eliminate the vision tower and causal decoder typical in visual document retrieval systems. The model processes multilingual text tokens and raw 32×32 RGB image patches through a unified Transformer, trained via masked diffusion denoising with cross-modal conditioning. NeoMME-Retriever achieves 0.523 nDCG@10 on ViDoRe v3 (260M), matching 3.75B-parameter models while being 14.4× smaller. Hierarchical token pooling and quantization reduce index size by 255.5× with minimal performance drop. Weaknesses include text-only retrieval and natural-image transfer.
multimodal encodermasked diffusionhierarchical token poolingsingle-tower architecturecross-modal conditioning
Meta FAIR Introduces AI Research Preference Models (RPMs): Ranking ML Experiments Before Spending GPU Hours
Meta FAIR introduces Research Preference Models (RPMs) to optimize GPU resource allocation in AI research by ranking unexecuted ML experiment candidates. RPMs employ frozen pretrained LLMs (Qwen3.6-27B) without fine-tuning, operating within the AIRA-dojo scaffold via pairwise knockout tournaments over 15 candidates. Two variants are proposed: (1) an inference-only judge (57.7–59.0% offline accuracy) and (2) an agentic judge running small-scale pilots (capped at 30 trials). On AIRS-Bench, RPMs improve average normalized scores from 0.684 (random) to 0.711 (inference) and 0.729 (agentic), achieving 1.5–1.6× speedup to baseline performance. New SOTA results include 94.1% on WinoGrande (agentic) and 95.7% on SVAMP (inference).
research preference modelsaira-dojoqwen3.6-27bairs-benchknockout tournament
UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents
UC Berkeley researchers introduce CUA-Lite, an open platform unifying sandboxes, datasets, evaluation, and reinforcement learning for computer-use agents (CUAs). The system integrates agents, environments, traces, and training under a single action space and LiteSample schema, addressing fragmentation in existing tools. Key innovations include Lite.OSWorld, a Docker-based alternative to QEMU/KVM VMs for OSWorld tasks, reducing memory usage from 4.1 GB to 0.9 GB while maintaining fidelity (matched scores across 13 models). The platform supports 30k+ tasks, 15+ benchmarks, and 10+ agents, with preprocessed datasets (e.g., Aguvis, OpenCUA) available on Hugging Face. Deployment requires only Docker, enabling cloud and CI compatibility.
computer-use agentslite.osworldlitesamplereinforcement learningdocker
Perplexity Details Its GPU Embedding Stack: How Ivy, Tulip and ROSE Serve pplx-embed
Perplexity Engineering details its GPU-optimized embedding stack for pplx-embed, comprising three services: Ivy (Rust HTTP gateway for CPU-side preprocessing), Tulip (gRPC inference server for scheduling), and ROSE (Python-based runtime with CUDA graph management). The system unifies batch (throughput-optimized) and online (latency-sensitive) embedding workloads by repurposing LLM prefill/decode kernels. Key optimizations include lazy CUDA graph capture (reducing startup overhead), LazyTensor async execution, and FlashInfer modifications enabling full-model graph capture. The architecture achieves efficiency by saturating GPU utilization at ~512 tokens per batch for sub-billion-parameter models, with linear latency scaling dominated by dense layers rather than attention.
embedding servingcuda graphslazytensorflashinfergrpc inference
MG Ship adds AI route optimisation as logistics returns accelerate
MG Ship introduced an AI route optimization module integrating dynamic routing algorithms and carrier scoring for global logistics. The system processes live transit data, weather patterns, port congestion, and carrier performance metrics to recommend low-cost, low-risk routes. Enterprise deployments report 15–20% fuel savings, 15–25% faster transit, and 12–22% lower costs, with payback in 3–6 months. Carrier scoring evaluates on-time performance, claims rates, and total cost-to-serve, reducing lead-time variance and expedited freight spend. Scenario simulations model peak shipping conditions, improving on-time-in-full delivery rates.
route optimizationcarrier scoringtransit reliabilitypredictive analyticssupply chain intelligence
Generated automatically at 2026-09-07 21:55 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.