ReCrit frames critic interaction as a correctness-transition problem and uses quadrant-based RL rewards to improve LLM performance on scientific reasoning benchmarks by rewarding corrections and robustness while penalizing sycophancy.
hub
Critique-grpo: Advancing llm reasoning with natural language and numerical feedback
18 Pith papers cite this work. Polarity classification is still indexing.
18
Pith papers citing it
abstract
Recent advances in reinforcement learning (RL) using numerical rewards have significantly enhanced the complex reasoning capabilities of large language models (LLMs). However, we identify three fundamental limitations of purely numerical feedback: performance plateaus, ineffective spontaneous self-reflection, and persistent failures. We show that plateaued RL models can successfully refine failed solutions when given natural language critiques. Motivated by this, we propose Critique-GRPO, an online RL framework that integrates both natural language and numerical feedback for policy optimization. This approach enables LLMs to learn simultaneously from initial responses and critique-guided refinements, effectively internalizing the exploration benefits of both stages. Extensive experiments show that Critique-GRPO outperforms all compared supervised and RL-based fine-tuning methods, achieving average Pass@1 improvements of approximately +15.0-21.6% on various Qwen models and +7.3% on Llama-3.2-3B-Instruct across eight challenging reasoning tasks. Notably, Critique-GRPO facilitates effective self-improvement through self-critiquing, achieving substantial gains over GRPO, e.g., a +16.7% Pass@1 improvement on AIME 2024. The code and models are released at: https://github.com/zhangxy-2019/critique-GRPO
hub tools
citation-role summary
background 3
citation-polarity summary
roles
background 3polarities
background 3representative citing papers
The cancellation hypothesis shows how rollout-level rewards produce token-level credit assignment in critic-free RL through cancellation of opposing signals on shared tokens, with empirical support and batching interventions that enhance performance.
Gen-Searcher is the first trained search-augmented image generation agent using SFT followed by GRPO reinforcement learning with dual text-image rewards, delivering 15-16 point gains on knowledge-intensive benchmarks.
Video-R1 uses temporal-aware RL and mixed datasets to boost video reasoning in MLLMs, with a 7B model reaching 37.1% on VSI-Bench and surpassing GPT-4o.
REVES augments LLM post-training by decoupling revision and verification signals from successful multi-step trajectories, reporting +6.5 point gains on LiveCodeBench over RL baselines.
The paper introduces Random-Reset Policy Optimization (RRPO) and Self-Reset Policy Optimization (SRPO) that use resets to enable more precise credit assignment in RL for language model reasoning, with SRPO outperforming GRPO and RRPO across benchmarks.
DITTO uses RL with verbal feedback to train LLMs for human behavior simulation, reporting 36% average gains over base models and outperforming GPT-5.4 on 6 of 10 SOUL benchmark tasks.
A two-stage probe (hidden-state estimate plus attention-based correction) yields per-step GRPO rewards that survive prefix contamination and beat external-judge and tree-search rewards in the reported benchmarks.
STRIDE co-trains generator and verifier on outcome rewards alone to deliver learnable stepwise language feedback that redirects LLM reasoning trajectories and outperforms scalar-reward baselines.
ICRL uses joint RL training of solver and critic with distribution-calibration re-weighting and role-wise advantage estimation to internalize critique into unassisted LLM performance, yielding 6.4-point gains on agentic tasks and 7.0 on math reasoning with Qwen3 models.
MOPD improves on-policy distillation by using peer successes and failures from multiple rollouts to construct more informative teacher signals, yielding consistent gains over baselines on reasoning benchmarks.
Span-level Wasserstein distances between hidden-state distributions of correct and incorrect rollouts provide a self-supervised signal to reweight advantages in GRPO, improving fine-grained credit assignment on math and code tasks.
AdaTooler-V trains MLLMs to adaptively use vision tools via AT-GRPO reinforcement learning and new datasets, reaching 89.8% on V* and outperforming GPT-4o.
TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.
Bi-NAC frames RL with textual feedback as a Stackelberg bilevel program and reports that 2B and 6B models trained this way outperform larger GRPO baselines on MATH-500 and GPQA.
Skill1 trains a single RL policy to co-evolve skill selection, utilization, and distillation in language model agents from one task-outcome reward, using low-frequency trends to credit selection and high-frequency variation to credit distillation, outperforming baselines on ALFWorld and WebShop.
LLM post-training is unified as off-policy or on-policy interventions that expand support for useful behaviors, reshape policies within reachable states, or consolidate behavior across training stages.
OneThinker unifies image and video reasoning in one model across 10 tasks via a 600k corpus, CoT-annotated SFT, and EMA-GRPO reinforcement learning, reporting strong results on 31 benchmarks plus some cross-task transfer.
citing papers explorer
-
ReCrit: Transition-Aware Reinforcement Learning for Scientific Critic Reasoning
ReCrit frames critic interaction as a correctness-transition problem and uses quadrant-based RL rewards to improve LLM performance on scientific reasoning benchmarks by rewarding corrections and robustness while penalizing sycophancy.
-
The Cancellation Hypothesis in Critic-Free RL: From Outcome Rewards to Token Credits
The cancellation hypothesis shows how rollout-level rewards produce token-level credit assignment in critic-free RL through cancellation of opposing signals on shared tokens, with empirical support and batching interventions that enhance performance.
-
Gen-Searcher: Reinforcing Agentic Search for Image Generation
Gen-Searcher is the first trained search-augmented image generation agent using SFT followed by GRPO reinforcement learning with dual text-image rewards, delivering 15-16 point gains on knowledge-intensive benchmarks.
-
Video-R1: Reinforcing Video Reasoning in MLLMs
Video-R1 uses temporal-aware RL and mixed datasets to boost video reasoning in MLLMs, with a 7B model reaching 37.1% on VSI-Bench and surpassing GPT-4o.
-
REVES: REvision and VErification--Augmented Training for Test-Time Scaling
REVES augments LLM post-training by decoupling revision and verification signals from successful multi-step trajectories, reporting +6.5 point gains on LiveCodeBench over RL baselines.
-
Credit Assignment with Resets in Language Model Reasoning
The paper introduces Random-Reset Policy Optimization (RRPO) and Self-Reset Policy Optimization (SRPO) that use resets to enable more precise credit assignment in RL for language model reasoning, with SRPO outperforming GRPO and RRPO across benchmarks.
-
Reinforcing Human Behavior Simulation via Verbal Feedback
DITTO uses RL with verbal feedback to train LLMs for human behavior simulation, reporting 36% average gains over base models and outperforming GPT-5.4 on 6 of 10 SOUL benchmark tasks.
-
PAIR: Prefix-Aware Internal Reward Model for Multi-Turn Agent Optimization
A two-stage probe (hidden-state estimate plus attention-based correction) yields per-step GRPO rewards that survive prefix contamination and beat external-judge and tree-search rewards in the reported benchmarks.
-
STRIDE: Learnable Stepwise Language Feedback for LLM Reasoning
STRIDE co-trains generator and verifier on outcome rewards alone to deliver learnable stepwise language feedback that redirects LLM reasoning trajectories and outperforms scalar-reward baselines.
-
ICRL: Learning to Internalize Self-Critique with Reinforcement Learning
ICRL uses joint RL training of solver and critic with distribution-calibration re-weighting and role-wise advantage estimation to internalize critique into unassisted LLM performance, yielding 6.4-point gains on agentic tasks and 7.0 on math reasoning with Qwen3 models.
-
Multi-Rollout On-Policy Distillation via Peer Successes and Failures
MOPD improves on-policy distillation by using peer successes and failures from multiple rollouts to construct more informative teacher signals, yielding consistent gains over baselines on reasoning benchmarks.
-
Hidden States Know Where Reasoning Diverges: Credit Assignment via Span-Level Wasserstein Distance
Span-level Wasserstein distances between hidden-state distributions of correct and incorrect rollouts provide a self-supervised signal to reweight advantages in GRPO, improving fine-grained credit assignment on math and code tasks.
-
AdaTooler-V: Adaptive Tool-Use for Images and Videos
AdaTooler-V trains MLLMs to adaptively use vision tools via AT-GRPO reinforcement learning and new datasets, reaching 89.8% on V* and outperforming GPT-4o.
-
Trust Region On-Policy Distillation
TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.
-
RL with Learnable Textual Feedback: A Bilevel Approach
Bi-NAC frames RL with textual feedback as a Stackelberg bilevel program and reports that 2B and 6B models trained this way outperform larger GRPO baselines on MATH-500 and GPQA.
-
Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning
Skill1 trains a single RL policy to co-evolve skill selection, utilization, and distillation in language model agents from one task-outcome reward, using low-frequency trends to credit selection and high-frequency variation to credit distillation, outperforming baselines on ALFWorld and WebShop.
-
Large Language Model Post-Training: A Unified View of Off-Policy and On-Policy Learning
LLM post-training is unified as off-policy or on-policy interventions that expand support for useful behaviors, reshape policies within reachable states, or consolidate behavior across training stages.
-
OneThinker: All-in-one Reasoning Model for Image and Video
OneThinker unifies image and video reasoning in one model across 10 tasks via a 600k corpus, CoT-annotated SFT, and EMA-GRPO reinforcement learning, reporting strong results on 31 benchmarks plus some cross-task transfer.