CEPO: RLVR Self-Distillation using Contrastive Evidence Policy Optimization

• URL: http://arxiv.org/abs/2605.19436v1
• Date: Tue, 19 May 2026 06:46:19 GMT
• Title: CEPO: RLVR Self-Distillation using Contrastive Evidence Policy Optimization
• Abstract summary: A model produces a correct solution under reinforcement learning with verifiable rewards (RLVR)<n>Every token receives the same reward signal regardless of whether it was a decisive reasoning step or a grammatical filler.<n>We propose Contrastive Evidence Policy Optimization (CEPO)<n>CEPO asks a sharper question at every token: "does the correct answer favor this token?" but "does the correct answer favor it while the wrong answer disfavors it?"
• Score: 50.59956036193097
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: When a model produces a correct solution under reinforcement learning with verifiable rewards (RLVR), every token receives the same reward signal regardless of whether it was a decisive reasoning step or a grammatical filler. A natural fix is to condition the model on the correct answer as a teacher, identifying tokens it would have generated differently had it known the answer. Prior work shows this either corrupts training by leaking the answer into the gradient, or produces a weak signal that cannot distinguish decisive steps from filler, since both look equally surprising relative to the model's baseline. We propose Contrastive Evidence Policy Optimization (CEPO), which asks a sharper question at every token: not just "does the correct answer favor this token?" but "does the correct answer favor it while the wrong answer disfavors it?" A token satisfying both is a genuine reasoning step; one satisfying neither is filler. The wrong-answer teacher is constructed from rejected rollouts already in the training batch, incurring no additional sampling cost. We prove CEPO inherits all structural safety guarantees of the prior state of the art while strictly sharpening credit at decisive tokens, with the improvement vanishing exactly at filler positions. Empirically, CEPO achieves 43.43% and 60.56% average accuracy across five multimodal mathematical reasoning benchmarks at 2B and 4B scale, respectively, versus 41.17% and 57.43% for GRPO under identical training budgets. Distribution-matching self-distillation methods (OPSD, SDPO) fall below the untrained baseline, empirically confirming the information leakage our theory predicts. Our code is available at https://github.com/ahmedheakl/CEPO.

Related papers

• TTPO: Test-Time Policy Optimization [53.81524570216593]
  Test-Time Policy Optimization (TTPO) is an asymmetric objective that distills agreeing rollouts via OPSD and penalizes disagreeing rollouts with Grouped RL.<n>We show that TTPO matches label-supervised OPSD on five competition-level benchmarks, raises Qwen3-1.7B from 38.0% to 45.2% in TTT, yields +25.2% to +36.4% without thinking, and shows strong cross-task generalization.
  arXiv  Detail & Related papers  (2026-08-27T17:58:10Z)
• Privileged, but Biased: How PI-Conditioned Teachers Break Self-Distillation [3.5592415157621105]
  Self-distillation (SD) has emerged as a compute-efficient alternative to reinforcement learning with verifiable rewards.<n>We reproduce SDPO's reported gains in its easy setting, then apply the identical setup to difficult tasks and find that it does not.<n>We explain this failure through a single causal chain from the loss to the model it produces.
  arXiv  Detail & Related papers  (2026-08-05T12:59:10Z)
• Reinforcement Learning for Large Language Model Selective Evidence Adoption from Contaminated Retrieval Results [3.9964880913101677]
  Retrieval-augmented large language models face contexts that interleave useful evidence with misleading statements or instruction-like content.<n>We introduce SelectBench, a benchmark and training set for selective evidence adoption, and post-train Qwen3.5-4B directly with DAPO.<n>On the corrected 325-example SelectBench-v2 test set, strict success rises from 22.46% for the original checkpoint to 25.54% with DAPO-Rule and 26.46% with DAPO-DeepSeek.
  arXiv  Detail & Related papers  (2026-07-22T12:45:04Z)
• H$^2$SD: Hybrid Hindsight Self-Distillation [61.71131241473028]
  Reinforcement learning with verifiable rewards (RLVR) provides reliable outcome supervision for language model reasoning.<n>Existing self-distillation methods add a privileged teacher but typically assign it a fixed role.<n>We introduce Hybrid Hindsight Self-Distillation ($mathrmH2mathrmSD$), which jointly adapts teacher context and update strategy to trajectory correctness.
  arXiv  Detail & Related papers  (2026-07-21T10:47:27Z)
• Localizing Credit at the Divergence: Path-Conditioned Self-Distillation for LLM Reasoning [12.363923974156636]
  Reinforcement learning from verifiable rewards assigns a single scalar to each rollout.<n>On-policy self-distillation addresses this by letting the same model act as a teacher conditioned on privileged information.<n>We propose Hindsight Self-Distillation (HSD), which conditions the teacher on a successful peer rollout.
  arXiv  Detail & Related papers  (2026-06-14T03:37:27Z)
• Right Makes Might: Aligning Verified Hidden States Empowers RL Reasoning [55.264863369127774]
  Current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states.<n>We propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference.<n>On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales.
  arXiv  Detail & Related papers  (2026-06-02T06:51:15Z)
• Pause and Reflect: Conformal Aggregation for Chain-of-Thought Reasoning [8.024041325202612]
  Chain-of-thought (CoT) reasoning with self-consistency improves performance by aggregating multiple sampled reasoning paths.<n>We introduce a conformal procedure for CoT reasoning that directly addresses aggregation uncertainty.<n>Our approach replaces majority voting with weighted score aggregation over reasoning paths and calibrates an abstention rule using conformal risk control.
  arXiv  Detail & Related papers  (2026-05-13T20:33:59Z)
• Unleashing Implicit Rewards: Prefix-Value Learning for Distribution-Level Optimization [74.91418266859297]
  Implicit process reward models (PRMs) provide fine-grained reward signals along the reasoning process.<n>PRMs suffer from a train-inference mismatch: training only constrains a sequence-level aggregate, whereas inference requires token-level scores to reflect local step quality.<n>We propose a novel Implicit Prefix-Value Reward Model (IPVRM), which directly learns a prefix-conditioned value function estimating the probability of eventual correctness.<n>We also propose Distribution-Level RL (DistRL), which computes TD advantages for both sampled tokens and high-probability candidate tokens.
  arXiv  Detail & Related papers  (2026-04-14T18:19:54Z)
• IAPO: Information-Aware Policy Optimization for Token-Efficient Reasoning [47.55414301744048]
  We argue that existing sequence-level reward-shaping methods offer limited control over how reasoning effort is allocated across tokens.<n>We propose IAPO, an information-theoretic post-training framework that assigns token-wise advantages based on each token's conditional mutual information.<n>IAPO consistently improves reasoning accuracy while reducing reasoning length by up to 36%, outperforming existing token-efficient RL methods.
  arXiv  Detail & Related papers  (2026-02-22T05:30:14Z)
• Mitigating Spurious Correlations Between Question and Answer via Chain-of-Thought Correctness Perception Distillation [25.195244084313114]
  Chain-of-Thought Correctness Perception Distillation (CoPeD) aims to improve the reasoning quality of the student model.<n>CoPeD encourages the student model to predict answers based on correct rationales and revise them when they are incorrect.
  arXiv  Detail & Related papers  (2025-09-06T05:33:17Z)
• The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning [37.13807960501503]
  Reinforcement learning with verifiable rewards (RLVR) is a promising approach for training language models (LMs)<n>We decompose the learning signal into reinforcing correct responses and penalizing incorrect ones, referred to as Positive and Negative Sample Reinforcement (PSR and NSR)<n>We show that NSR works by suppressing incorrect generations and redistributing probability mass toward other plausible candidates, guided by the model's prior beliefs.
  arXiv  Detail & Related papers  (2025-06-02T06:10:54Z)
• Exploring the Limit of Outcome Reward for Learning Mathematical Reasoning [65.2421542320293]
  Reasoning abilities are crucial components of general intelligence.<n>Recent advances by proprietary companies, such as o-series models of OpenAI, have made remarkable progress on reasoning tasks.<n>This paper proposes a new RL framework, termed OREAL, to pursue the performance limit that can be achieved through textbfOutcome textbfREwtextbfArd-based reinforcement textbfLearning for mathematical reasoning tasks.
  arXiv  Detail & Related papers  (2025-02-10T18:57:29Z)
• From $r$ to $Q^*$: Your Language Model is Secretly a Q-Function [50.812404038684505]
  We show that we can derive DPO in the token-level MDP as a general inverse Q-learning algorithm, which satisfies the Bellman equation. We discuss applications of our work, including information elicitation in multi-turn dialogue, reasoning, agentic applications and end-to-end training of multi-model systems.
  arXiv  Detail & Related papers  (2024-04-18T17:37:02Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.