Metraipolis › archive › dc_8fafdf24931e

I-SDPO: Instance-Level Adaptive Self-Distillation Policy Optimization — arXiv:2608.12957

Source
arXiv (en)
Original
https://arxiv.org/abs/2608.12957
License (as archived)
CC0-1.0
Archived
2026-10-19T04:24:41.687Z
SHA-256 of the archived text
0b0eb45a4c782113f2fea9cbcfbb27d4b8dd1d471ce767d5b7fa0fb30c4b2841
Stable URL of this page
https://metraipolis.com/documento/dc_8fafdf24931e

How to cite

I-SDPO: Instance-Level Adaptive Self-Distillation Policy Optimization — arXiv:2608.12957 — arXiv (en), https://arxiv.org/abs/2608.12957 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_8fafdf24931e (Metraipolis archive dc_8fafdf24931e, sha256 0b0eb45a4c782113f2fea9cbcfbb27d4b8dd1d471ce767d5b7fa0fb30c4b2841)
BibTeX
@misc{dc8fafdf24931e,
  title = {I-SDPO: Instance-Level Adaptive Self-Distillation Policy Optimization — arXiv:2608.12957},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2608.12957},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_8fafdf24931e; sha256 0b0eb45a4c782113f2fea9cbcfbb27d4b8dd1d471ce767d5b7fa0fb30c4b2841},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.12957. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Group Relative Policy Optimization (GRPO) learns from reward differences within a rollout group, but receives no useful relative signal when every sampled response is incorrect. Privileged self-distillation can fill this gap with dense token supervision, yet applying it throughout training creates a different failure mode: the teacher is a biased, low-variance surrogate for the reward objective, so persistent imitation can oppose reward-improving updates after the policy becomes capable of producing successful trajectories. We introduce I-SDPO (Instance-Level Adaptive Self-Distillation Policy Optimization), which treats teacher reliance as capability-dependent. I-SDPO makes one routing decision per input instance and shares it across that instance's rollout group: all-incorrect groups use a privileged self-distillation objective, whereas any-success groups remain intact for GRPO. This design uses imitation only where group-relative rewards are uninformative. A local analysis characterizes when teacher and reward directions align and shows that a non-vanishing biased distillation weight induces an optimization bias floor. The routing rule automatically reduces the expected distillation rate as success probability rises, withdrawing teacher influence without a hand-designed schedule. On SciKnowEval, I-SDPO obtains the best result in all four scientific domains and improves average mean@16 accuracy from 56.67% with GRPO to 70.31%, with a maximum domain gain of 18.24 points.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-14 in cs.CL).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2608.12957 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-19T04:17:42.411Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-08-14?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).