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.
Figma shares how its latest unreleased editing model, post-trained on Qwen 3.8 27B, scores 77.8% on internal evals, surpassing the top third-party model, while the general availability agent sees 20% faster task completion.
RL environments generate infinite training signals by letting agents learn through experience, shifting post-training costs from per-task human labeling to one-time environment construction.
Confirmed that removing the KL term is essential for preventing performance collapse and improving efficiency when training Qwen3 0.6B with GRPO.
The model surpasses leading open-weight models on MindCube spatial reasoning and outperforms other world models on FlowBench motion prediction.
The framework integrates model serving, trajectory generation, and policy updates into a single reproducible pipeline supporting Qwen and GLM models.
The number of tasks where Qwen3.8-27B achieved 20 consecutive successes increased from 38 to 107.
The 3T-class open-weight model supports a 1M token context window and reduces claim-level hallucination rates from 23.4% to 8.3%.
The method, presented at ICML 2026, generates observation and action models via LLMs to reconfigure the agent-environment interface, outperforming human-designed baselines in locomotion and manipulation tasks.
Standard instruction-following rewards cause models to collapse to short, empty answers; LIFT-Bench fixes this by scoring content alongside formatting constraints.
NVIDIA has released Nemotron 3.5 Bolt E4B, a small model for interactive agents that operates on 4GB VRAM.
OpenEnv joined the AAIF project and released v0.10.0, enabling agent learning environment development to proceed under a neutral foundation.
The paper demonstrates that N individuals acting as parallel environment rollouts create an effective per-sample step size of alpha = 1/N.
The method uses executable database state checks as binary rewards to fine-tune agents, reducing the cost per dependable task from $24.36 to $8.00 for Qwen3.8-27B.
NVIDIA, Pollen Robotics, and Hugging Face introduce a workflow to train walking policies for the Microduck robot in Isaac Lab 3.0 and deploy them via ONNX.
The A5X platform integrates Vera Rubin NVL72 with Google Cloud Virgo Network to support large-scale reinforcement learning and agentic inference.
JOLT jointly trains a single model as both teacher and student to provide dense, aligned guidance, achieving higher accuracy and significantly better sample efficiency than GRPO on math, coding, and agentic benchmarks.
Fine-tuned on 2B tokens of synthetic data, the model predicts screen deltas from keypresses, outperforming prompted baselines on shell and file readback tasks.
The model jointly optimizes task generation, scaffold construction, and solution rollouts via reinforcement learning to replace fixed human-curated training sets.
HybridAgent achieved a 10.4% performance improvement and cost reduction, while Raw-RAG outperformed complex memory systems.
The update enables Multi-Token Prediction for the 12B-parameter Mellum2.1 Thinking model, which features 2.5B active parameters and a 131k context window.
Long agentic rollouts, such as those in Alibaba's QwenGyre paper lasting over four hours, exacerbate on-policy drift and require complex solutions like tree serialization and bitwise-consistent kernels.
The framework improves reasoning accuracy in 1.5B parameter models by integrating In-Context Learning into Test-Time Reinforcement Learning.
Microsoft Research has released StudentSim, a simulator that mimics individual students' learning patterns, and demonstrated how to use it to validate the instructional quality of AI tutors.
The expected gradient is a covariance and the Fisher is a variance, capping updates at |G| ≤ std(A) · √F.