Amazon Web Services detailed Amazon SageMaker AI multi-turn reinforcement learning (MTRL), a fine-tuning capability that trains LLM search agents across full multi-turn trajectories rather than single…
A 28 Sep 2026 arXiv paper introduces ProVer, a framework that targets potentially pivotal decisions for fine-grained credit assignment in agentic reinforcement learning, addressing what the authors ca…
GraphCert, a post-training method from researchers behind the arXiv paper 2609.38798v1, bootstraps agentic graph reasoning by generating graph-grounded QA pairs and certifying supporting evidence into…
A working implementation of reinforcement learning for calibrated decisions (RLCD) for reasoning models, OpenJev-RLCD, matches or beats supervised fine-tuning, RFT/STaR and GRPO in accuracy and beats …
A new method called RLTL;DR lets a Qwen 3.5 9B Thinking policy break through a learning barrier on tool-calling and coding datasets filtered to Pass@128 = 0, reaching a Pass@1 of 14–31% with insights …
A new reinforcement learning technique called Off-Policy-Aware Cross-Model Trajectory Exchange targets the all-fail group problem in Reinforcement Learning with Verifiable Rewards (RLVR) methods such …
Alibaba's Qwen team released Qwen-Audio-3.1, a five-model audio stack led by the full-duplex speech model Qwen-Audio-3.1-Realtime, available as a managed API on QwenCloud over WebSocket with no open w…
A new method called SLCA-GRPO addresses cross-segment credit misattribution in tool-calling reinforcement learning, targeting a structural failure mode where standard on-policy RL algorithms such as G…
GEPA, a prompt-optimization method in which a model reads an agent's full trace — reasoning, tool calls and errors — and rewrites the prompt, reportedly doubled the gains of GRPO after a single round …
Xiaomi's MiMo-V2.6 models use groupwise agentic grading to replace binary pass/fail rewards in reinforcement learning, ranking multiple rollouts per task against each other to preserve gradient signal…
A paper titled "Learning to Solve Hard Problems in RL for LLMs by Never Giving Up" proposes countering the Matthew Effect in reinforcement learning, where LLMs improve far more on easy problems they a…
A developer published a practical guide to GRPO (Group Relative Policy Optimization), an RL method for fine-tuning and aligning large language models that avoids the separate critic/value model requir…
Researchers introduced Dual-Entropy Enhanced Policy Optimization (DEEPO), a dual-stage reinforcement learning method that reduces hallucination in multimodal large language models by combining semanti…
A new arXiv paper (arXiv:2609.27035v1) introduces Reinforcement Learning with Decomposed Subtasks (RLDS), a method that replaces the scalar GRPO advantage with Subtask-Decomposed Advantage Estimation …
Researchers introduced RECAP (REdundancy-aware Credit Assignment via Propagation), a method that assigns step-level credit in large reasoning models using an LLM-annotated semantic dependency graph an…
A latent-GRPO training step for the Qwen 3.8 27B model dropped from a median of 36 minutes on a single H100 to 2.6 minutes on one 8× NVIDIA B300 node, a 14× speedup, according to the training team's a…
A new arXiv paper (2609.20846v1) reports that fine-tuning several 4B large reasoning models with a novel GRPO reward that encourages efficient reasoning about whether a task contains enough informatio…
A multimodal reasoning framework using Task-Aware Group Relative Policy Optimization (GRPO) achieved an Overall Score of 83.24 on the official PCBA Standard-to-Real Grand Challenge leaderboard, accord…
COAL-SQL, a unified post-training framework combining Coverage-Guided Augmentation (CGA) and Failure-Driven Learning (FDL), achieves 64.9% execution accuracy on the BIRD development set using only 12,…
Kuaishou deployed OneBid, a unified auto-bidding foundation model built on Decision Transformer, reporting an overall +2.2% ADVV gain on oCPX Ads and a peak +13.1% gain in the ROAS scenario in online …