Metraipolis › archive › dc_e05069d1eb84

Teach Yourself Where to Look: On-Policy Attention Self-Distillation for Reasoning — arXiv:2609.33200

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.33200
License (as archived)
CC0-1.0
Archived
2026-10-07T21:53:59.968Z
SHA-256 of the archived text
448d585e5dd805da02931779c17aa3c862b936e036044b5a60fc5a795c3e1383
Stable URL of this page
https://metraipolis.com/documento/dc_e05069d1eb84

How to cite

Teach Yourself Where to Look: On-Policy Attention Self-Distillation for Reasoning — arXiv:2609.33200 — arXiv (en), https://arxiv.org/abs/2609.33200 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_e05069d1eb84 (Metraipolis archive dc_e05069d1eb84, sha256 448d585e5dd805da02931779c17aa3c862b936e036044b5a60fc5a795c3e1383)
BibTeX
@misc{dce05069d1eb84,
  title = {Teach Yourself Where to Look: On-Policy Attention Self-Distillation for Reasoning — arXiv:2609.33200},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.33200},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_e05069d1eb84; sha256 448d585e5dd805da02931779c17aa3c862b936e036044b5a60fc5a795c3e1383},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.33200. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

On-policy self-distillation trains reasoning models on their own trajectories using dense token distribution guidance from a privileged teacher with access to a verified solution. This supervision transfers what the teacher predicts without directly transferring where it attends within the preceding context. We introduce On-Policy Attention Self-Distillation (OPASD), which complements token-level supervision with solution-conditioned attention distillation. Because the privileged teacher can attend to verified solution tokens unavailable to the student, OPASD projects teacher attention onto student-visible positions and renormalizes the resulting distribution before alignment. Across three model sizes and four competition-level mathematics benchmarks, OPASD consistently outperforms token-only OPSD, improving average accuracy by 4.98 to 8.40 percentage points. OPASD also avoids the response-length inflation and performance degradation observed with token-only distillation, reducing generated rollout tokens by 73.9% and estimated model compute by 72.6% while training 1.53x faster. These results show that solution-conditioned attention provides a complementary supervision signal that makes on-policy self-distillation more accurate, stable, and compute-efficient.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-29 in cs.CL).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.33200 — 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 2026-10-07T21:27:08.753Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-29?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).