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.
AesCode achieves an overall score of 85.89 on infographic generation benchmarks, outperforming GPT-5.5 and Claude Opus 4.8 by using decoupled cross-modal rewards.
StudentSim outperforms baselines like GPT-5.4 and Maia2 in chess, writing, and math simulations, achieving 90.5% tutor accuracy in reinforcement learning tests.
The framework achieves up to 1.85x speedup over Colocate and lifts NL2RepoBench scores from 52.5% to 58.5% on Qwen3.8-2.4T-A95B.
Microsoft Research Asia released Agent Lightning v1.0, a 3,500-line framework that raised Qwen3.5-9B's Pass@1 on SWE-bench Verified by 14.6 percentage points using only 6,000 training samples.
Nemotron-3-Ultra-CC scored 535.4/600 at IOI 2026, surpassing the top human score, while a specialized IMO system achieved 30/42 points.
Microsoft Research scientists Jennifer Neville and Chad Atalla discuss how AI models degrade in multi-turn interactions despite high single-turn benchmark scores, highlighting the need for new evaluation methods and user strategies.
The model features 49 billion active parameters and will have its weights released by the end of October 2026.
RISED uses LLM-generated rubrics to guide data selection and self-distillation, achieving the highest mean pass rate across diverse interactive environments.
Reflection AI has introduced Beam, a 501B-parameter open-weight MoE model trained on 10.5K NVIDIA GB300 GPUs, claiming high inference efficiency compared to larger models like GLM 5.2 and Kimi K3.
The 4B-parameter model is derived from Qwen3.5-4B and trained on 1,500 synthetic SWE tasks using the Leaf harness.
ProVer achieves up to 9.91% relative improvement over standard GRPO on agentic tasks by verifying pivotal decision segments instead of evaluating every step.
The AI, trained on just 16 GPUs for a fraction of DeepMind’s cost, won 15 of 20 matches against four-time world champion Pim Niemeijer.
Clef models are open-sourced under Apache 2.0 on Hugging Face and hosted on Workers AI, claiming to outperform Jev on the Decision Index.
Apple researchers introduce RLTL;DR, a method that boosts Pass@1 scores from near-zero to 12–13% on difficult tasks by internalizing self-generated feedback.
OpenAI shares initial guidelines for safety cases in frontier reinforcement learning training, outlining recommended technical safeguards, operational approvals, and incident investigation protocols.
This article derives policy gradient methods for LLMs, explaining how REINFORCE evolves into GRPO through baseline variance reduction and addressing the gradient bias caused by inference engine mismatches.
The new MOPD-based solution reduced repetition rates significantly at a cost of approximately $90,000, compared to an estimated $2.31 million for a full MixRL restart.
The DSec platform supports over 380,000 concurrent sandboxes and 5,000 creations per second to enable efficient reinforcement learning for LLM agents.
Trajectory introduces Intelligence Density and density-aware training, demonstrating significant reductions in output tokens and improved efficiency on legal and financial benchmarks without sacrificing accuracy.
The guide details how to use the managed RLFT service to adapt Gemini for tasks that are easy to score but hard to demonstrate, such as code execution and content moderation.
The model uses a single reinforcement-learning stage without prior domain-specific supervised fine-tuning.
Stanford researchers propose EXPO-FT, an algorithm for universal post-training in robotics that achieves 100% success rates on physical tasks by using a lightweight editing policy to stabilize reinforcement learning.
A simulation replicating the entire fruit fly brain has been revealed, demonstrating its ability to learn Super Mario Bros. 3 in real time.
Google Research and UIUC unveiled R4T, which distills RL-trained query fan-out behaviors into a 53.9M parameter diffusion model, reducing search latency by 12–20x.