H-OPD：置信度感知的异构多教师多模态同策略蒸馏框架

View PDF HTML (experimental)

Abstract:On-policy distillation (OPD) has recently emerged as an effective post-training paradigm by providing supervision on student-generated trajectories. However, existing OPD methods for multimodal reasoning usually rely on a static teacher routing, assigning each sample to a single teacher based on modality or task type. This ignores that visual grounding and abstract reasoning may dominate different decoding steps, making a single teacher insufficient for the full trajectory. To this end, H-OPD is proposed as a confidence-aware heterogeneous multi-teacher OPD framework for multimodal reasoning. By verifying the complementarity of heterogeneous teachers in the same reasoning process, H-OPD replaces task or sample level teacher routing with token-level teacher arbitration along the shared student trajectory. H-OPD employs vision-to-language description transfer to enable text-only teachers to access key visual semantics, and uses a confidence-aware arbitration mechanism to dynamically combine vision-language teacher and text-only teachers at each token. Extensive evaluations over 11 widely-used reasoning benchmarks showcase the superior performance of our method.

Comments:	EMNLP2026
Subjects:	Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
Cite as:	arXiv:2607.02592 [cs.CV]
	(or arXiv:2607.02592v3 [cs.CV] for this version)
	https://doi.org/10.48550/arXiv.2607.02592 arXiv-issued DOI via DataCite

Submission history

From: Qixiang Yin [view email]
[v1] Wed, 1 Jul 2026 07:37:10 UTC (3,164 KB)
[v2] Mon, 24 Aug 2026 07:42:13 UTC (3,167 KB)
[v3] Fri, 9 Oct 2026 08:17:54 UTC (3,167 KB)