Vision-OPD: On-Policy Self-Distillation Improves Fine-Grained Vision in Multimodal LLMs

Forum topic · 小凯 · 2026-05-20

Summary

Researchers Qianhao Yuan, Jie Lou, and Xing Yu propose Vision-OPD (Vision On-Policy Distillation), a regional-to-global self-distillation framework for multimodal large language models (MLLMs), presented in arXiv paper 2505.14302. The work starts from an observed regional-to-global perception gap: the same MLLM answers fine-grained questions more accurately when conditioned on evidence-centered crops than on the corresponding full images, indicating that many failures come from difficulty focusing on relevant evidence rather than weak local recognition. Vision-OPD instantiates two conditional policies from a single MLLM: a crop-conditioned teacher and a full-image-conditioned student. The student generates on-policy rollouts, and the method minimizes token-level divergence between teacher and student next-token distributions over those rollouts. This transfers the model's own privileged regional perception into its full-image policy, internalizing the benefit of visual zooming without external teacher models, ground-truth labels, reward verifiers, or inference-time tool use. Experiments on multiple fine-grained visual understanding benchmarks show Vision-OPD models match or outperform larger open-source, closed-source, and 'Thinking-with-Images' agentic models.

Paper Overview

Field: NLP Authors: Qianhao Yuan, Jie Lou, Xing Yu Published: 2026-05-19 arXiv: 2505.14302

Key Points

Multimodal Large Language Models (MLLMs) still struggle with fine-grained visual understanding, where answers often depend on small but decisive evidence within the full image.
The authors observe a regional-to-global perception gap: the same MLLM answers fine-grained questions more accurately when conditioned on evidence-centered crops than on the corresponding full images. This suggests many failures stem from difficulty focusing on relevant evidence, not insufficient local recognition ability.
Vision-OPD (Vision On-Policy Distillation) is a regional-to-global self-distillation framework that transfers the model's own privileged regional perception to its full-image policy.

Method

Vision-OPD instantiates two conditional policies from the same MLLM instance:
A crop-conditioned teacher
A full-image-conditioned student
The student generates on-policy rollouts, and Vision-OPD minimizes the token-level divergence between the teacher's and student's next-token distributions over those rollouts.
This lets the model internalize the benefit of visual zooming without external teacher models, ground-truth labels, reward verifiers, or inference-time tool use.

Results

On multiple fine-grained visual understanding benchmarks, Vision-OPD models achieve performance comparable to or better than larger open-source, closed-source, and 'Thinking-with-Images' agentic models.

Original paper: https://arxiv.org/abs/2505.14302