ROSD: Reflective On-Policy Self-Distillation for Language Model Reasoning across Domains

• URL: http://arxiv.org/abs/2605.28014v1
• Date: Wed, 27 May 2026 06:09:29 GMT
• Title: ROSD: Reflective On-Policy Self-Distillation for Language Model Reasoning across Domains
• Abstract summary: On-policy self-distillation (OPSD) improves the reasoning performance of large language models (LLMs)<n>Existing OPSD methods often yield limited gains on in-domain reasoning and generalize poorly to out-of-domain problems.<n>We propose Reflective On-policy Self-Distillation (ROSD), a framework that turns reference-solution imitation into targeted reasoning correction.
• Score: 40.558397280256685
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: On-policy self-distillation (OPSD) improves the reasoning performance of large language models (LLMs) by providing dense token-level supervision for on-policy rollouts. However, existing OPSD methods often yield limited gains on in-domain reasoning and generalize poorly to out-of-domain problems. We identify two key causes: conditioning the self-teacher on a verified solution encourages imitation of training-domain reference trajectories rather than error-specific correction, and applying distillation to the full response can overwrite valid reasoning prefixes and reinforce overfitting. We propose Reflective On-policy Self-Distillation (ROSD), a framework that turns reference-solution imitation into targeted reasoning correction through reflection-guided, error-localized distillation. For each rollout, ROSD uses a self-reflector to extract a corrective idea and locate the first erroneous span. The corrective idea guides the self-teacher toward targeted supervision, while the localized error span restricts distillation to where correction is needed. This design corrects flawed reasoning while preserving valid prefixes. Experiments on multiple in-domain and out-of-domain reasoning benchmarks show that ROSD yields stronger in-domain reasoning performance overall and substantially better out-of-domain generalization than standard OPSD. Code is available at https://github.com/ZiqiZhao1/ROSD.

Related papers

• dOPSD: On-Policy Self-Distillation for Diffusion Language Models [52.60302464420127]
  Diffusion large language models generate text by iteratively denoising a masked sequence.<n>Supervised fine-tuning is off-policy and suffers from exposure bias, while reinforcement learning gives only sparse, sequence-level rewards.<n>On-policy self-distillation (OPSD) offers a promising alternative, using one model as both student and teacher.
  arXiv  Detail & Related papers  (2026-07-05T17:47:29Z)
• Purified OPSD: On-Policy Self-Distillation Without Losing How to Think [52.879387744920415]
  We find that on-policy self-distillation (OPSD) consistently fails on long chain-of-thought (long-CoT) reasoning models.<n>We propose a two-step solution: first, we construct a reference-only teacher to isolate the non-transferable component of the supervision signal.<n>Second, we use pointwise mutual information (PMI) as the mechanism to transform this residual into a well-formed PMI target distribution.
  arXiv  Detail & Related papers  (2026-07-02T14:33:07Z)
• Rethinking Reward Supervision: Rubric-Conditioned Self-Distillation [60.55792673956761]
  We propose textbfRubric-Conditioned Self-Distillation, a framework that incorporates rubrics as structured, fine-grained feedback for on-policy self-distillation.<n>We evaluate on a diverse suite of science reasoning benchmarks and results show that rubric-conditioned self-distillation effectively converts rubric-level criteria into token-level guidance.
  arXiv  Detail & Related papers  (2026-06-17T17:54:04Z)
• Trajectory-Refined Distillation [25.346810464266497]
  Trajectory-Refined Distillation (TRD) is a trajectory-level correction method that revises the student's rollout under the teacher guidance while within on-policy support.<n> TRD can also be applied to on-policy self-distillation (OPSD), a parameter-sharing variant that uses the student model conditioned on privileged informations as the teacher.
  arXiv  Detail & Related papers  (2026-06-07T03:17:25Z)
• AMR-SD: Asymmetric Meta-Reflective Self-Distillation for Token-Level Credit Assignment [39.63424981516754]
  Asymmetric Meta-Reflective Self-Distillation (AMR-SD)<n>We introduce Causal Information Gain (CIG) with an asymmetric, ReLU-gated threshold to translate these reflections into sparse, highly precise token-level advantage modulations.<n> Experiments across scientific, mathematical, and tool-use benchmarks demonstrate that AMR-SD significantly outperforms existing baselines.
  arXiv  Detail & Related papers  (2026-05-18T15:14:34Z)
• Validity-Calibrated Reasoning Distillation [4.277471273303775]
  Reasoning distillation aims to transfer multi-step reasoning capabilities from large language models to smaller, more efficient ones.<n>Recent methods rely on static teacher-student hierarchies and frame distillation as trajectory imitation.<n>We propose validity-calibrated reasoning distillation, a framework that treats reasoning distillation as a problem of local learning-signal allocation.
  arXiv  Detail & Related papers  (2026-04-14T12:32:12Z)
• Reforming the Mechanism: Editing Reasoning Patterns in LLMs with Circuit Reshaping [51.33479921152361]
  Large language models (LLMs) often exhibit flawed reasoning ability that undermines reliability.<n>Existing approaches to improving reasoning typically treat it as a general and monolithic skill, applying broad training which is inefficient and unable to target specific reasoning errors.<n>We introduce Reasoning Editing, a paradigm for selectively modifying specific reasoning patterns in LLMs while preserving other reasoning pathways.
  arXiv  Detail & Related papers  (2026-03-06T22:51:39Z)
• APR: Penalizing Structural Redundancy in Large Reasoning Models via Anchor-based Process Rewards [61.52322047892064]
  Test-Time Scaling (TTS) has significantly enhanced the capabilities of Large Reasoning Models (LRMs)<n>We observe that LRMs frequently conduct repetitive self-verification without revision even after obtaining the final answer during the reasoning process.<n>We propose Anchor-based Process Reward (APR), a structure-aware reward shaping method that localizes the reasoning anchor and penalizes exclusively the post-anchor AST.
  arXiv  Detail & Related papers  (2026-01-31T14:53:20Z)
• Save the Good Prefix: Precise Error Penalization via Process-Supervised RL to Enhance LLM Reasoning [59.76691952347156]
  Reinforcement learning (RL) has emerged as a powerful framework for improving the reasoning capabilities of large language models (LLMs)<n>Most existing RL approaches rely on sparse outcome rewards, which fail to credit correct intermediate steps in partially successful solutions.<n>We propose Verifiable Prefix Policy Optimization (VPPO), which uses PRMs only to localize the first error during RL.
  arXiv  Detail & Related papers  (2026-01-26T21:38:20Z)
• Directional Attractors in LLM Reasoning: How Similarity Retrieval Steers Iterative Summarization Based Reasoning [0.0]
  We introduce InftyThink with Cross-Chain Memory, an extension that augments iterative reasoning with an embedding-based semantic cache of previously successful reasoning patterns.<n> Experiments show that semantic lemma retrieval improves accuracy in structured domains while exposing failure modes in tests that include heterogeneous domains.
  arXiv  Detail & Related papers  (2025-12-22T00:26:54Z)
• In-Token Rationality Optimization: Towards Accurate and Concise LLM Reasoning via Self-Feedback [38.915062716409686]
  InTRO is a new framework that enables both token-level exploration and self-feedback for accurate and concise reasoning.<n>InTRO consistently outperforms other baselines, raising solution accuracy by up to 20% relative to the base model.<n>Its chains of thought are notably more concise, exhibiting reduced verbosity.
  arXiv  Detail & Related papers  (2025-11-13T01:47:06Z)
• Retrieval is Not Enough: Enhancing RAG Reasoning through Test-Time Critique and Optimization [58.390885294401066]
  Retrieval-augmented generation (RAG) has become a widely adopted paradigm for enabling knowledge-grounded large language models (LLMs)<n>RAG pipelines often fail to ensure that model reasoning remains consistent with the evidence retrieved, leading to factual inconsistencies or unsupported conclusions.<n>We propose AlignRAG, a novel iterative framework grounded in Critique-Driven Alignment (CDA)<n>We introduce AlignRAG-auto, an autonomous variant that dynamically terminates refinement, removing the need to pre-specify the number of critique iterations.
  arXiv  Detail & Related papers  (2025-04-21T04:56:47Z)

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.