Metraipolis › archive › dc_9dfac8928ee6
Edit, Don't Imitate: Minimal-Edit On-Policy Distillation for Multimodal Reasoning — arXiv:2610.30231
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2610.30231
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-25T08:33:47.736Z
- SHA-256 of the archived text
f9203e99cc4c70034077a093fc6a3463a09387d5437813954696b2bd17d6d71a- Stable URL of this page
https://metraipolis.com/documento/dc_9dfac8928ee6
How to cite
Edit, Don't Imitate: Minimal-Edit On-Policy Distillation for Multimodal Reasoning — arXiv:2610.30231 — arXiv (en), https://arxiv.org/abs/2610.30231 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_9dfac8928ee6 (Metraipolis archive dc_9dfac8928ee6, sha256 f9203e99cc4c70034077a093fc6a3463a09387d5437813954696b2bd17d6d71a)
BibTeX
@misc{dc9dfac8928ee6,
title = {Edit, Don't Imitate: Minimal-Edit On-Policy Distillation for Multimodal Reasoning — arXiv:2610.30231},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2610.30231},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_9dfac8928ee6; sha256 f9203e99cc4c70034077a093fc6a3463a09387d5437813954696b2bd17d6d71a},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.30231. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Visual mathematical reasoning can fail before reasoning begins: when a vision-language model (VLM) misreads a diagram, subsequent reasoning proceeds from false premises. This perceptual bottleneck is especially pronounced in smaller models, while accurate textual descriptions can substantially narrow the gap to larger models. Because obtaining such descriptions at inference requires additional annotation or model calls, we propose minimal-edit corrective on-policy distillation. Using reference figure descriptions and verified reference solutions as training-only information, a teacher scores incorrect student rollouts under a minimal-edit objective, and only teacher-disfavored tokens contribute to the corrective loss. Correct rollouts receive task-reward reinforcement learning. This preserves on-policy supervision while avoiding full-response imitation. Across model scales, the method requires no descriptions at test time and meets or exceeds the corresponding base model supplied with a reference figure description at inference. It also improves perception-oriented performance, particularly low-level geometric perception, and transfers to external mathematics and perception benchmarks. Because the minimal-edit teacher evaluates corrections rather than generating its own solution, the method remains effective with a frozen same-size teacher, although its gains are smaller than with a larger teacher on most metrics. Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno). Categorie arXiv: Machine Learning (cs.LG) ; Artificial Intelligence (cs.AI). Pagina del lavoro: https://arxiv.org/abs/2610.30231 — 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 2026-10-25T08:31:13.129Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).