PROTEA is an interface for offline debugging and refinement of multi-agent LLM workflows. It evaluates intermediate outputs with configurable rubrics, localizes bottlenecks via workflow graph visualization, and generates targeted prompt revisions. On two production-adjacent workflows, PROTEA improves document-inspection accuracy from 64.3% to 83.9% and recommendation Hit@5 from 0.30 to 0.38.
Vision-OPD introduces regional-to-global self-distillation to improve fine-grained visual understanding in MLLMs. The framework transfers the model's privileged perception on evidence-centered crops to its full-image policy via token-level KL divergence minimization on on-policy rollouts. Competitive results on fine-grained visual understanding benchmarks without external models or ground-truth labels.
ReTAMamba proposes a Mamba-based architecture for predicting irregular clinical time series. The model estimates observation reliability from missingness and elapsed time, integrates short/long-term information via Chronological Weaving, and uses a budgeted token router. On MIMIC-IV, eICU, and PhysioNet 2012, AUPRC gains of 7.51%, 7.80%, and 10.15% respectively.
TabTrim, a novel table pruning framework for TableQA, replaces sequential revisions with gold trajectory-supervised parallel search. The system uses intermediate sub-tables from gold SQL queries to train a pruner and verifier. TabTrim-8B achieves 73.5% average accuracy, outperforming strongest baselines by 3.2% (79.4% on WikiTQ, 61.2% on TableBench).
QuickLAP fuses physical and language feedback to learn robot reward functions in real time using a Bayesian framework. LLMs extract reward feature attention masks and preference shifts from free-form utterances, integrated with physical corrections via closed-form update rule. Achieves 70% error reduction vs physical-only and heuristic multimodal baselines in semi-autonomous driving simulator.
ALIGN is a vision-language framework to infer precise accident coordinates from Bangla news reports and map-based cues. Using an agentic architecture combining OCR, LLM, and vision-language models, the system reduces localization error from 10.9 km to 0.593 km on validation data and 0.465 km on official Dhaka Metropolitan Police records.
Custom CUDA runtime for small-batch inference (robotics, VLA, world models). Bottlenecks are not GEMM alone but runtime overhead: kernel fragmentation, layout transitions, precision conversions (FP8/FP4), Python scheduling. Results: Pi0.5 on RTX 5090 ~17.6ms, GROOT N1.6 ~12.5-13.1ms, Qwen 27B ~129 tok/s.
Sub-JEPA improves LeWorldModel (LeCun's group, NYU) by applying Gaussian regularization within frozen random orthogonal subspaces instead of globally. Gains up to +10.7 pp on Two-Room, straighter latent trajectories, better physical state decodability. Code and paper released.
GitBook hosts 30,000 documentation sites on Vercel, serving 120 million monthly page views. The platform uses Next.js `use cache` directive to invalidate cache in under 300ms per site, processing 40,000 daily invalidations. 41% of traffic comes from AI crawlers.
OpenAI and Snowflake announce a $200M partnership to embed frontier AI models directly into Snowflake's data platform. Customers can deploy AI agents and extract insights without data movement. Native integration of OpenAI models within Snowflake's ecosystem.
OpenAI and SoftBank Group partner via SB Energy to build multi-gigawatt AI data center campuses, including a 1.2 GW Texas facility supporting the Stargate initiative.
OpenAI releases GPT-5.2, its strongest model for math and science, achieving state-of-the-art results on GPQA Diamond and FrontierMath benchmarks. The model solves an open theoretical problem and generates reliable mathematical proofs.
OpenAI releases GPT-5.1 in API with faster adaptive reasoning, extended prompt caching, improved coding performance, and new apply_patch and shell tools.
Meta releases Llama 3.2 with native vision capabilities and device-optimized versions. The model processes images and text natively, available in 1B and 3B variants for on-device execution.
Google releases Gemma 2 2B, a lightweight model optimized for inference. ShieldGemma provides protection against harmful content. Gemma Scope offers interpretability tools to analyze model internals.
Hugging Face's Transformers Code Agent achieves 92% accuracy on the GAIA benchmark, outperforming Claude 3.5 Sonnet (92%) and GPT-4o (87.9%). The agent combines web search, code execution, and multi-step reasoning to solve complex tasks.
Hugging Face releases Idefics2, an open-source 8B vision-language model handling images and text. Supports up to 1024×1024 pixel resolution and runs on standard hardware. Available under Apache 2.0 license with public weights and code.
Hugging Face releases StarCoder2, open-source code model trained on The Stack v2 (17B code tokens dataset). StarCoder2 outperforms CodeLlama on HumanEval and MBPP benchmarks. Available in 3B, 7B, and 15B variants with open weights and source code.
OpenAI explores leveraging deep learning's generalization properties to control strong models with weak supervisors. New research direction for superalignment with promising initial results.
OpenAI trains a neural network to play Minecraft using Video PreTraining (VPT) on unlabeled human gameplay videos. The model learns to craft diamond tools (24,000-action task) with minimal labeled data. It uses native human interface (keyboard/mouse) and represents progress toward general computer-using agents.
Vercel introduces 'The Agent Stack', a complete framework for building production-grade AI agents. It combines AI SDK (unified multi-model interface), AI Gateway (centralized routing and billing), and enables calling Claude, GPT and others without vendor lock-in.
Xcientist is a research harness that externalizes research synthesis and experimental validation for AI scientists into inspectable, contract-governed processes. It organizes literature evidence, idea states, implementation plans, and repair traces as persistent research artifacts, eliminating claim drift where runnable artifacts no longer support the originally claimed mechanism.
SGCD, an iterative self-improvement framework, addresses off-trajectory states in GUI agents. The system first runs a plain policy, then uses a skill-guided policy to generate successful continuations. On OSWorld-Verified, SGCD improves success rates of three base models from ~30% to over 50%.
Causal framework for sleep recovery scoring from multimodal polysomnography. Uses DAG learning on two cohorts (MESA n=1540, MrOS n=825) to identify five physiological domains (respiratory burden, hypoxia, fragmentation, architecture, autonomic regulation). Sleep Recovery Score (SRS) achieves 2.5× stronger alignment with perceived recovery than standard AHI.
LLMZero uses LLM agents with tree search to discover adaptive RL training strategies. The system identifies that capacity parameters accumulate monotonically while regularization parameters oscillate. Across 4 GRPO tasks, discovered strategies outperform the base model by 9-140% and grid search by 6-15%.
ASTRA is an air traffic control training simulator automating pilot roles through speech recognition, instruction interpretation, and response generation. The system reduces Word Error Rate from 107.80% to 23.45% on Singaporean-accented aviation speech, and evaluates trainee radiotelephony communications achieving 91.7% accuracy, 88.2% brevity, and 86.9% completeness scores.
SAGE is a post-hoc method to improve selective unlearning in LLMs. It corrects final update vectors by suppressing components damaging retention, without rerunning the original unlearning pipeline. Tested across multiple methods and scales, SAGE reduces the forget-retain trade-off.
Mechanistic analysis of mathematical reasoning in multilingual LLMs. Math-associated parameters exhibit partial cross-lingual overlap, concentrated in intermediate layers. English produces the largest set of math-relevant parameters, while lower-resource languages reveal smaller parameter sets.
Safety Reflection Pretraining inserts short safety reflections into pretraining corpora to establish self-monitoring directly in language modeling. On 1.7B models pretrained on FineWeb-Edu, the method improves safety classification accuracy and substantially reduces success rates of inference-stage and finetuning attacks.
Framework for customization and efficient deployment of LLM-based multi-agent systems in enterprise settings. Combines continual pretraining, supervised fine-tuning, and preference optimization to adapt compact models to specialized domains. Integrates speculative decoding and FP8 quantization to reduce latency and costs. Achieves 4.48x throughput speedup while maintaining performance.
Novel LLM personalization: store user facts as surgical edits in a hash-keyed memory table (Engram) instead of global LoRA. Reduces memory footprint by 33,000x, improves indirect-reasoning accuracy by 5.6x on average, and enables stacking multiple users without cross-contamination.
Amazon, Nvidia, and AMD invest $310 million in Odyssey ML, a 3D world model startup valued at $1.45 billion. IQT fund and Google's Jeff Dean join the round. World models are emerging as the next major AI bet after language models.
Zhipu AI releases GLM-5.2 under MIT license with stable 1-million-token context. On FrontierSWE benchmark for long-duration coding tasks, the open-source model trails Anthropic's Claude Opus 4.8 by just one percentage point. Significant gap remains on reasoning versus closed-source rivals.
Gemma 4 E2B runs in-browser at 255 tokens/sec using WebGPU kernels optimized by Fable 5. Demo and kernels released on Hugging Face.
Vercel unveils agent-first infrastructure at Ship 2026 in London. Three core components: Agent Stack (building blocks for agents), Vercel Connect (secure external tool access without persistent tokens), and eve (open-source framework for production agents with durable execution, sandboxed compute, approvals, and evals).
Researchers from Nvidia, Carnegie Mellon University, and UC Berkeley use AI coding agents to teach robots dexterous grasping in real-world conditions. A fleet of eight robots achieves 99% success rate on complex tasks.
DeepSeek closes a funding round exceeding $7 billion, among the largest in the AI sector. Record amount for the Chinese startup specializing in language models.
High-performance code intelligence MCP server. Indexes codebases into persistent knowledge graph in milliseconds. Supports 158 languages, sub-ms queries, 99% fewer tokens. Single static binary, zero dependencies.
TimesFM is a pretrained foundation model developed by Google Research for time-series forecasting. The GitHub repository provides an open-source implementation of this specialized model.
ByteDance releases UI-TARS-desktop, an open-source multimodal AI agent stack connecting cutting-edge AI models and agent infrastructure. Platform for building agents capable of interacting with user interfaces.