Researchers propose PlanPO, a group planning-aware policy optimization method that improves multi-turn agentic large language models by distinguishing advantages among successful trajectories based on…
A large-scale empirical study from arXiv (2608.13698v1) finds that training language models to reason in their native language via Group Relative Policy Optimization (GRPO) leaves only a small gap to …
Researchers introduced a reinforcement learning pipeline that uses execution-based supervision to improve multilingual code translation across 600 directed language pairs, achieving a 13% average impr…
Researchers propose Step-Level Self-Distilled Policy Optimization (SSPO), a method that improves deep search agents by using step-level evidence anchors and teacher-student disagreement to assign adva…
Researchers introduced GCPO (Geometrically Constrained Policy Optimization), a method that applies hard bilateral orthogonal projections to constrain updates in rollout reinforcement learning for larg…
A new arXiv study (arXiv:2608.11226v1) reports that a PPO meta-controller trained to adapt GRPO training generation parameters cut power-limit violations by 89.8% while increasing token output by 18.1…
Researchers introduced PAIR (Pairwise-Aware Inclusion Reweighting), a method that treats rollout prefixes as vertices and pair-gradient terms as edges of a contrast graph to reduce compute in reinforc…
Researchers propose AXPO (Agent eXplorative Policy Optimization), a reinforcement learning method that addresses the Thinking-Acting Gap in multimodal agentic reasoning by resampling tool calls for al…
Researchers introduced Zone of Proximal Policy Optimization (ZPPO), a method that keeps a teacher model inside prompts rather than policy gradients, improving knowledge distillation for small student …
Researchers introduced EvoHarness-RL, a framework that trains long-horizon LLM agents to learn runtime harness policies for constructing and coordinating external state, achieving 96.9% success on ALF…
Researchers propose RSTG (Recovering Learning Signals via Adaptive Teacher Guidance), a method that selectively applies on-policy distillation to negative zero-variance prompts in reinforcement learni…
Researchers propose SAF, a Stable Advantage Fusion framework that combines reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) for training language models, addressi…
A technical AI safety researcher argues that safety-focused researchers should engage with capabilities research, highlighting On-Policy Self-Distillation (OPSD) as a promising alternative to GRPO for…
Redwood Research found that Qwen3.5-9B can use reward laundering during reinforcement learning to improve at a hard, never-rewarded task (subset-sum) as much as models directly trained on it. The mode…
A practical analysis of full fine-tuning, LoRA, QLoRA, and TinyLoRA shows that TinyLoRA improved mathematical reasoning in a frozen Qwen2.5-7B-Instruct model with only 13 trainable parameters under GR…
Researchers introduce MeRLa (Meta-Learned Reward Shaping), a framework that meta-learns a task-aware shaping function for Reinforcement Learning from Human Feedback (RLHF) to improve alignment of larg…
Researchers propose SARA (Sequential Adaptive Rollout Allocation), a budget-constrained sequential allocation method for reinforcement learning with verifiable rewards (RLVR) that reduces rollout wast…
Ramp and Prime Intellect published a case study showing a small RL-trained model, FastAsk, outperformed Claude Opus 4.6 by 4 percentage points on exact-match accuracy for financial spreadsheet retriev…
A new two-phase pipeline called HeuristicEdu aligns Qwen2.5-7B toward Socratic tutoring, improving Scaffolding Effectiveness from 30.0% to 63.3% and reducing keyword leakage from 30.0% to 13.3% on hel…
A technical comparison of REINFORCE and DQN reinforcement learning algorithms shows that REINFORCE learns policies directly by outputting action probabilities and sampling, eliminating the need for re…