A new arXiv study (2607.20543v1) finds that reinforcement learning with verifiable rewards (RLVR) can improve one-sample accuracy while causing pass@k inversion, where the trained policy solves fewer …
A new study reveals that LUNAR, a state-of-the-art LLM unlearning method, fails to truly remove knowledge, as two attack routes can recover "forgotten" information. The author demonstrates that LUNAR'…
The llm-d project introduces co-operative time-slicing to interleave independent reinforcement learning jobs on shared hardware, increasing aggregate accelerator duty cycles from ~40% to 70% without i…
Google's Tunix post-training library introduces a high-throughput framework for agentic reinforcement learning that keeps TPUs fully utilized during multi-turn training. Tunix uses an asynchronous tra…
Researchers propose Group Entropy-Controlled Policy Optimization (GEPO), a lightweight extension to GRPO that uses group entropy to perform entropy-conditioned asymmetric advantage shaping, addressing…
Researchers propose W2SPO, an off-policy reinforcement learning method that uses short auxiliary segments from a weaker model to improve reasoning in large language models, achieving a Pass@1 improvem…
Researchers at arXiv propose Contextual Information-Gain Policy Optimization (CIGPO) to fix a zero-advantage lock-in failure in outcome-only reinforcement learning for multi-turn evidence-reading agen…
An independent researcher seeking an arXiv endorsement for cs.LG reports that a causal study of latent reasoning in a chess-playing LLM found that reinforcement learning (GRPO) improved legal-move rat…
Researchers propose Process-Scorer Guided Adaptive Tree Rollout (PATR), a quality-aware rollout framework for multi-turn reinforcement learning (RL) that uses process feedback to selectively branch fr…
A new method called Object-Part Hierarchical Reflective Grounding (OP-HRG) improves part-level visual grounding in multimodal large language models by introducing a coarse-to-fine reasoning pipeline t…
A developer evaluating the SDAR paper's claims found that the reported +9.4% accuracy gain on ALFWorld is less interesting than SDAR's ability to avoid training instability. The developer designed a t…
New research shows that incentivizing diversity in AI exploration through a representation-based bonus derived from hidden states of pre-trained models boosts verifier efficiency by over 50% for the Q…
ECHO, a novel framework for traceable context reconstruction in agentic reinforcement learning, achieves 43.4% held-out accuracy on BrowseComp-Plus, outperforming GRPO's 28.9% and SUPO's 36.1% with fe…
Researchers introduced LLM-as-a-Verifier, a general-purpose verification framework that computes continuous scores from scoring token logits to determine solution correctness without additional traini…
An open-source pipeline called ai-trains-ai uses nested reinforcement learning loops where an outer agent is RL-trained to write inner training jobs for small models, with the entire system run for ro…
Researchers have introduced STAMP (Step-wise Attribution of Modulated Potential), a new reinforcement learning approach that addresses the reward-credit mismatch by linking actions to rewards more dir…
A new framework called Self-Verified Reasoner (SVR-R1) allows AI models to self-verify their answers by asking themselves 'Yes or No' before finalizing a response, boosting accuracy on vision-language…
Agentic-DPO, a new method for training language models using offline preference supervision without costly environment rollouts, improved accuracy on tau-bench from 21.7% to 41.4% for a 9B model, matc…
Researchers introduce GEPA (Genetic-Pareto), a prompt optimizer that uses natural language reflection to learn from trial and error, outperforming reinforcement learning methods like GRPO by 6% on ave…
A study on training reasoning language models (RLMs) in Japanese, led by researchers using Qwen-3-Swallow-8B, finds that reasoning-language control is feasible but achieves performance at best on par …