{"paper":{"title":"Learning with Rare Success but Rich Feedback via Reflection-Enhanced Self-Distillation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reflection-Enhanced Self-Distillation lets models learn from failure feedback by creating diagnostic reflections and a reusable global playbook.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bing Yin, Changlong Yu, Chengyu Dong, Haoran Liu, Ilgee Hong, Jingbo Shang, Qin Lu, Sha Li, Shuowei Jin, Xintong Li, Yuwei Zhang, Zhenyu Shi","submitted_at":"2026-05-12T20:46:05Z","abstract_excerpt":"Enabling Large Language Models (LLMs) to continuously improve from environmental interactions is a central challenge in post-training. While on-policy self-distillation offers a promising paradigm, existing methods predominantly treat environmental feedback as a passive conditioning signal. Consequently, they heavily rely on successful demonstrations and struggle to learn in rare-success regimes. To bridge this gap, we introduce Reflection-Enhanced Self-Distillation (RESD), a framework that transforms raw failure feedback into an active source of corrective supervision. Instead of passively ap"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"RESD substantially outperforms standard self-distillation baselines and achieves significantly faster early-stage improvement than GRPO with 8× samples using only a single rollout per prompt.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the model-generated retrospective reflections accurately diagnose local errors and that the curated global playbook preserves reusable lessons without introducing noise or compounding errors across training steps.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RESD turns failure trajectories into token-level supervision via retrospective reflections and a persistent global playbook, enabling faster improvement than standard self-distillation or GRPO with only one rollout per prompt.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reflection-Enhanced Self-Distillation lets models learn from failure feedback by creating diagnostic reflections and a reusable global playbook.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"8daeae4eeeb4395ca66bc16ddc1f418291aa8157d8709b88316242c5e4caca0a"},"source":{"id":"2605.12741","kind":"arxiv","version":1},"verdict":{"id":"1bc594d4-bc48-4675-a39f-5986b6c1d8ee","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T21:20:07.083787Z","strongest_claim":"RESD substantially outperforms standard self-distillation baselines and achieves significantly faster early-stage improvement than GRPO with 8× samples using only a single rollout per prompt.","one_line_summary":"RESD turns failure trajectories into token-level supervision via retrospective reflections and a persistent global playbook, enabling faster improvement than standard self-distillation or GRPO with only one rollout per prompt.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the model-generated retrospective reflections accurately diagnose local errors and that the curated global playbook preserves reusable lessons without introducing noise or compounding errors across training steps.","pith_extraction_headline":"Reflection-Enhanced Self-Distillation lets models learn from failure feedback by creating diagnostic reflections and a reusable global playbook."},"references":{"count":32,"sample":[{"doi":"","year":2023,"title":"GKD: Generalized knowledge distillation for auto-regressive se- quence models.arXiv preprint arXiv:2306.13649","work_id":"78b7a553-fd43-4995-b559-e95779797d3f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"On-policy distillation of language models: Learning from self-generated mistakes","work_id":"3733ff2d-cd95-4776-9fa9-1b2328326749","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Retaining by doing: The role of on-policy data in mitigating forgetting, 2025","work_id":"4d844db6-2d57-4fe8-ace2-574a47636b4a","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models","work_id":"6b7f0773-4e99-4274-9d8e-279a1f25c5e1","ref_index":4,"cited_arxiv_id":"2401.01335","is_internal_anchor":true},{"doi":"","year":2026,"title":"Deepseek-v4: Towards highly efficient million-token context intelligence","work_id":"b5a5843d-47c0-432e-88d5-c9eaac555a7f","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":32,"snapshot_sha256":"baac3e7471575d1284ac586b57d1d22460486b0c63fd6a5bc820423b0f2c6c1c","internal_anchors":13},"formal_canon":{"evidence_count":1,"snapshot_sha256":"87f5a5f78cca71786cd7005c46a6a050286f8ebb09ae3b3d620050bf441c370b"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}