H-OPD: Confidence Aware Heterogeneous Multi-Teacher Multimodal On-policy Distillation
Quick Answer
H-OPD introduces a confidence-aware heterogeneous multi-teacher framework for on-policy distillation in multimodal reasoning, enhancing performance by dynamically combining teachers at the token level.
Quick Take
H-OPD introduces a confidence-aware heterogeneous multi-teacher framework for on-policy distillation in multimodal reasoning, enhancing performance by dynamically combining teachers at the token level. Extensive evaluations across 11 benchmarks demonstrate its superior capabilities compared to existing methods, addressing the limitations of static teacher routing.
Key Points
- H-OPD replaces static teacher routing with token-level arbitration for improved reasoning.
- Employs vision-to-language transfer for text-only teachers to access visual semantics.
- Demonstrated superior performance on 11 widely-used reasoning benchmarks.
- Dynamic combination of teachers enhances the quality of student-generated trajectories.
- Addresses limitations of existing on-policy distillation methods in multimodal contexts.
Paper Resources
📖 Reader Mode
~2 min readAbstract: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: | Technical report(work-in-progress). Code will be available at \url{this https URL} |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG) |
| Cite as: | arXiv:2607.02592 [cs.CV] |
| (or arXiv:2607.02592v1 [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)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CV
See more →LLM-Guided ANN Index Optimization for Human-Object Interaction Retrieval
A phase-aware LLM agent optimizes human-object interaction retrieval, outperforming Optuna TPE by 33.3% and VDTuner by 34.2% on the HICO-DET benchmark. This method enhances throughput by 15.3x over UniIR and demonstrates strong transferability across vector database management systems.