Read original ↗
paperarXivTrust 82 · PrimaryPublished yesterdayLive · 19h ago

Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

On-policy self-distillation (OPSD) has emerged as a promising paradigm for improving LLM reasoning, where a privileged teacher with access to reference solutions provides token-level supervision on the student's own generated trajectories. However, we find that OPSD consistently fails on long chain-of-thought (long-CoT) reasoning models, yielding at best marginal gains while destabilizing the reflective reasoning capability these models depend on. Through a novel decomposition of the teacher's supervision signal, we identify the root cause: the teacher's supervision is dominated by a reference

Lineage graph

Paper → model → repo connections mined from source citations (Tier-1 exact match).

Why these links exist

  • Linked via arxiv authorZhanming Shen

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorJintao Tong

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorShaotian Yan

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorChen Shen

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorHao Chen

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorWentao Ye

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorXiaomeng Hu

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorRui Miao

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorHaobo Wang

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorJunbo Zhao

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorGang Chen

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

  • Linked via arxiv authorJieping Ye

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

Implements

authored (incoming)

Implements (incoming)

Related across the graph

Topics