Title: Vision-OPD: Enhancing Multimodal LLMs’ Fine-Detail Perception Through On-Policy Self-Distillation

Abstract:

Multimodal Large Language Models (MLLMs) continue to face significant challenges in fine-grained visual comprehension, often requiring minute yet critical evidence from an entire image to formulate accurate responses. Our analysis reveals a distinct "regional-to-global" perception gap: an MLLM demonstrates higher accuracy on fine-grained queries when provided with evidence-centered image crops compared to the complete, unaltered image. This discrepancy indicates that many errors arise not from a lack of local recognition capabilities, but from an inability to effectively focus on relevant visual evidence.

To address this issue, we introduce Vision-OPD (Vision On-Policy Distillation), a self-distillation framework designed to bridge the regional-to-global divide by transferring the model’s superior regional perception into its full-image processing policy. Vision-OPD establishes two conditional policies derived from a single MLLM: a teacher conditioned on image crops and a student conditioned on full images. During training, the student produces on-policy rollouts, and the framework reduces token-level divergence between the next-token probability distributions of the teacher and student across these sequences.

This approach allows the model to learn the advantages of visual zooming internally, eliminating the need for external teacher models, ground-truth labels, reward verifiers, or inference-time tool usage. Evaluations across various fine-grained visual understanding benchmarks demonstrate that models trained with Vision-OPD deliver performance that is either competitive with or superior to significantly larger open-source and closed-source systems, as well as "Thinking-with-Images" agentic models. The source code is publicly accessible at https://github.com/VisionOPD/Vision-OPD.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC