OPD-V: Visual On-Policy Self-Distillation with Modality Balance

2026-08-05

AI Summary

This work addresses the limitation of existing online policy self-distillation methods in multimodal large language models, which often underutilize visual information due to text dominance during reasoning. To mitigate this issue, the authors propose a modality-balanced visual online policy self-distillation framework that treats modality balance as privileged information for the first time. Specifically, they construct positive and negative teacher models using Zoom-In and Masked images, and introduce a positive modality-balanced logits margin to define a confidence region for selecting reliable policy tokens to guide distillation. The proposed method consistently achieves significant performance gains across six benchmarks, four multimodal backbone architectures, and five post-training configurations, while simultaneously reducing training costs.

Abstract

On-Policy Self-Distillation (OPSD) has become a standard post-training approach for improving visual reasoning in multimodal large language models (MLLMs). Existing methods draw privileged information from diverse input sources to guide self-distillation. Yet these designs overlook Modality Imbalance, a challenge inherent to MLLM reasoning. When textual information dominates generation, the model cannot fully integrate its multimodal input. Consequently, carefully designed privileged information remains underused, limiting the effectiveness of OPSD. To examine this limitation, we construct a Positive Teacher with the Zoom-In Image and a Negative Teacher with the Mask Image, which exhibit different degrees of Modality Imbalance. Changes in their reasoning correctness and token logits reveal that Modality Balance can itself serve as privileged information. Motivated by this finding, we introduce OPD-V, a visual OPSD paradigm that instantiates such information through the Positive Teacher and Negative Teacher. Positive Modality-Balance Logits Margins define a Modality-Balance Trust Region that selects the on-policy tokens used for self-distillation. Experiments across 6 benchmarks, 4 MLLM backbones, and 5 post-training methods show that OPD-V consistently improves reasoning performance while reducing training cost.

Problem

• Modality Imbalance
• On-Policy Self-Distillation
• Multimodal Large Language Models
• Visual Reasoning
• Privileged Information

Innovation

• Modality Balance
• On-Policy Self-Distillation
• Visual Reasoning
• Multimodal Large Language Models
• Privileged Information