AI Briefing#reinforcement-learning
The latest AI and developer news about #reinforcement-learning, with the original source and a short summary.
The latest AI and developer news about #reinforcement-learning, with the original source and a short summary.
The November 2026 update introduces elastic gateway scaling, HTTP-based worker task pulling, and a shared Slime-based training pipeline for mathematical reasoning.
Optimization reduces planning time from 0.6–0.8 s to 45 ms, enabling 20 Hz camera processing on a single board.
The study shows that best-of-N alignment closely approaches the theoretical maximum reward gain, while PPO and GRPO remain substantially suboptimal.
The method learns a binary mask over gate projections, resulting in checkpoints hundreds to thousands of times smaller than full fine-tuning.
A steering vector derived from the difference between loose and strict reward checkpoints allows extrapolation to reduce resource-limit failures from 14.5% to 8.0%.
The Reshard operation announced by NVIDIA reduces weight transfer time in disaggregated RL environments by up to 5.7x.
NVIDIA releases GR00T 2, a foundation world-action model that predicts future video and robot actions, enabling zero-shot transfer from human demos and real-time control of Unitree G1 humanoids via the GR00T 2 Turbo variant.
Ai2 (AllenAI) has released ToyBlocks, an experimentation environment capable of analyzing the causes of robotic failures and generating data.
The framework explains that RL training amortizes search costs into model weights, making single samples likely to succeed without guaranteeing that any specific chain of thought is faithful or correct.
The RL update improves hard math and code scores to match the base Qwen3.8-27B while maintaining 0.67x thinking token usage on agent tasks.
The 50-page report details architecture, training, and evaluation for Drex models, which achieve 93.4% accuracy on long-context benchmarks.
The open-source framework pairs automatic data generation with benchmarks showing VLA models generalize poorly while frontier agents solve 30/30 stacking tasks.
Thinking Machines AI introduces Inkling 2 Small, a 594B MoE model with 17B active parameters, initially available with closed weights on Tinker, with full Apache 2.0 open weights planned for next week.
The 300B-scale model achieves 46.2% on Terminal-Bench 4.0 using only 64 NVIDIA B300 GPUs over five days.
The 276B-A12B model achieves 20.0% on Terminal-Bench 4.0 and 24.0% on FrontierCode 2.0, running at 200 tokens per second.
Axion 1.0 is a fully open agent foundation model trained in three stages: two lightweight SFT stages followed by reinforcement learning.
The system uses an Experience Flywheel and staged learning to close the loop between experience production and policy improvement.
The recipe uses time limits and shared rewards to incentivize parallel work, with Avalon-M1 collectives reaching nearly 2x the average useful work by step 200.
The open-source Cosmos Framework supports training and simulation for autonomous vehicles, robotics, and surgical workflows, while NIM enables sub-second live stream alerts.
The 1.6T parameter model achieves 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp, outperforming LongCat-2.0 and approaching frontier models like Claude-Opus-5.
A new theory paper demonstrates that under strategy specialization, every pass@1 improvement in RLVR strictly lowers success on some prompts, with a tight bound on the pass@k crossover.
The summary introduces three major trends—expanding agent evaluation scope, structured learning, and refining data quality—along with 10 papers.
The technique uses lossless GPU compression and Blackwell Decompression Engines to sync weights faster than intra-cluster P2P RDMA.
The framework decouples planning and grounding agents, using dense rewards and PPO to outperform Gemini-3.5-Flash on real iOS applications.