Research map / Large language models / Reasoning and chain of thought

Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Co-rewarding, a label-free self-supervised RL framework for LLM reasoning that uses cross-view agreement (rephrased questions) and an EMA-updated teacher to provide stable pseudo-label rewards, avoiding training collapse seen in self-rewarding methods.

Problem

Self-rewarding RL methods (self-certainty, entropy, majority voting) for LLM reasoning without labels often suffer training collapse and reward hacking, because single-view self-generated supervision is easily gamed.

Approach

Seeks complementary supervision across views: Co-rewarding-I uses cross-supervision between original and LLM-rephrased questions via majority-vote pseudo labels; Co-rewarding-II uses a slowly updated EMA teacher (the GRPO reference model) to supply pseudo labels decoupled from the online policy; Co-rewarding-III combines both.

Key results

• Co-rewarding outperforms self-rewarding baselines by +3.31% average on math benchmarks, with +7.49% on Llama-3.2-3B-Instruct.
• Co-rewarding-II reaches 94.01% Pass@1 on GSM8K with Qwen3-8B-Base trained on DAPO-14k, above its GT-Reward counterpart.
• Baselines (Self-Certainty, Entropy, Majority-Voting) collapse during training on MATH and DAPO-14k, while Co-rewarding-II stays stable across datasets.

Surprising findings

• Self-supervised Co-rewarding reaches or exceeds GT-reward RLVR on some benchmarks such as GSM8K and CRUX.
• Co-rewarding-I remains stable on MATH but still collapses on DAPO-14k, suggesting stability depends on training data.

Limitations

• Co-rewarding-I's stability depends on the training data; it collapses on DAPO-14k.
• Co-rewarding-I relies on Qwen3-32B for question rephrasing, adding an external model dependency.
• Gains over GT-reward are inconsistent and mostly on easier benchmarks such as GSM8K.

Experimental setup

Datasets: MATH DAPO-14k OpenRS MATH500 GSM8K AMC AIME24 LiveCodeBench CRUX IFEval MMLU-Pro

Benchmarks: MATH500 GSM8K AMC AIME24 LiveCodeBench CRUX IFEval MMLU-Pro

Metrics: pass@1 avg@8

Baselines: GT-Reward Self-Certainty Entropy Majority-Voting

Builds on: GRPO RLVR TTRL DAPO VeRL

Closest papers by meaning

• Consistent Paths Lead to Truth: Self-Rewarding Reinforcement Learning for LLM Reasoning (NeurIPS 2025 · similarity 0.93)
• LaSeR: Reinforcement Learning with Last-Token Self-Rewarding (ICLR 2026 · similarity 0.91)
• RM-R1: Reward Modeling as Reasoning (ICLR 2026 · similarity 0.91)
• Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs (ICLR 2026 · similarity 0.91)
• Self-Aligned Reward: Towards Effective and Efficient Reasoners (ICLR 2026 · similarity 0.91)
• Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling (NeurIPS 2025 · similarity 0.91)