Source-linked AI summary
Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation
Qianhao Yuan, Jie Lou, Xing Yu, Hongyu Lin, Le Sun, Xianpei Han, Yaojie Lu
TL;DR
MLLMs struggle to use small, decisive evidence embedded in full images, even when the same model succeeds on evidence-centered crops. Vision-OPD trains a full-image policy to match a crop-conditioned policy on its own rollouts, and achieves strong fine-grained understanding against much larger and agentic models.
Problem
MLLMs struggle with fine-grained visual understanding because small decisive evidence can be overwhelmed by global image context, despite being recognizable in regional crops.
Method
Vision-OPD uses a crop-conditioned teacher and full-image-conditioned student from the same MLLM, minimizing token-level distribution divergence along student on-policy rollouts.
Results
Vision-OPD substantially improves fine-grained understanding, surpassing much larger open-source, closed-source, and “Thinking-with-Images” agentic models across supported evaluations.
Takeaways & Limitations
The model’s privileged regional perception can supervise full-image inference without external teachers, ground-truth labels, reward verifiers, or inference-time visual tools.
Abstract
from arXiv · showhide
Multimodal Large Language Models (MLLMs) still struggle with fine-grained visual understanding, where answers often depend on small but decisive evidence in the full image. We 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, suggesting that many failures stem from difficulty to focus on relevant evidence rather than insufficient local recognition ability. Motivated by this observation, we propose Vision-OPD (Vision On-Policy Distillation), a regional-to-global self-distillation framework that transfers the model's own privileged regional perception to its full-image policy. Vision-OPD instantiates two conditional policies from the same MLLM: a crop-conditioned teacher and a full-image-conditioned student. The student generates on-policy rollouts, and Vision-OPD minimizes token-level divergence between the teacher and student next-token distributions along these rollouts. This enables the model to internalize 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 that Vision-OPD models achieve competitive or superior performance against much larger open-source, closed-source, and "Thinking-with-Images" agentic models. The code is available at https://github.com/VisionOPD/Vision-OPD
1 Introduction
MLLMs often miss small decisive details in full images despite recognizing them in evidence-centered crops. Vision-OPD transfers this privileged regional behavior to full-image inference through on-policy self-distillation, and experiments show substantial gains and a narrowed regional-to-global gap.
- Motivation: Fine-grained questions often depend on small visual details that are easy to overlook amid full-image visual tokens.This can produce plausible global-scene answers that miss the local evidence required by the question.
- Motivation: Evidence-centered crops often improve fine-grained accuracy because they make relevant evidence salient within the image.The same MLLM can answer more accurately from a crop than from the corresponding full image.
- Vision-OPD: Vision-OPD uses a crop-conditioned teacher and a full-image-conditioned student from the same MLLM, matching their token distributions on student-generated rollouts.The method transfers privileged regional behavior to the full-image policy without external teachers, labels, verifiers, or inference-time visual tools.
- Results: With only 6.2K synthetic training examples, Vision-OPD enables a 9B model to outperform much larger open-source and closed-source models and agentic visual-tool methods.On hold-out tasks beyond the training distribution, it maintains general visual understanding and reasoning ability.
- Results: Vision-OPD substantially narrows the regional-to-global gap, with on-policy sampling and dense supervision identified as important to its success.Comprehensive experiments validate the framework’s effectiveness.
2 Preliminary: distillation and on-policy distillation (OPD)
Traditional supervised distillation provides dense token-level guidance but trains on teacher-induced prefixes, creating a mismatch with the student’s inference states. OPD addresses this by sampling prefixes from the student and querying the teacher on those on-policy trajectories.
- Distillation: Traditional supervised distillation matches teacher distributions on teacher-induced prefixes, making it an off-policy training procedure.The student is trained on output sequences associated with supervised data rather than its own generated states.
- Distillation: Dense supervised distillation can create state-distribution mismatch because inference conditions on the student’s own prefixes, allowing errors to compound over long horizons.The training prefixes differ from those encountered during autoregressive generation.
- On-Policy Distillation: OPD samples output sequences from the current student and queries the teacher on prefixes generated by that student.This aligns teacher supervision with the student’s on-policy generation trajectory.
- On-Policy Distillation: OPD combines on-policy relevance from reinforcement learning with dense token-level guidance from distillation for autoregressive generation and reasoning.The combination supplies trajectory-relevant supervision without reducing guidance to only sequence-level feedback.
3 Vision-OPD
Vision-OPD addresses the regional-to-global perception gap by transferring a model’s crop-conditioned behavior to full-image inference through on-policy token-level self-distillation. It uses automatically synthesized crop–image–question triplets and matches teacher and student distributions along student-generated responses.
- Motivation: less is more for fine-grained visual understanding: Regional-input accuracy exceeds global-input accuracy by 18–22 points on ZoomBench, revealing that decisive evidence is often recognized in isolation but overlooked in full images.This gap persists across models, including larger and closed-source systems.
- Data synthesis: The synthesized dataset pairs each full image with an evidence-centered crop and a fine-grained question generated from retained small regions.Object identification and segmentation propose bounding boxes, and questions are generated to be answerable from the isolated region.
- Regional-to-global on-policy self-distillation: Vision-OPD instantiates a crop-conditioned teacher and full-image-conditioned student from the same MLLM to transfer privileged regional perception.The teacher sees the evidence-centered crop, while the student sees the full image and question.
- Regional-to-global on-policy self-distillation: The student samples on-policy responses, and training minimizes per-token divergence between teacher and student next-token distributions along the same student-generated prefixes.Gradients flow only through the student, aligning training states with inference and providing dense supervision.
- Design desiderata: Vision-OPD requires no external teacher, ground-truth labels, verifier, or inference-time tool use, and targets a single forward pass for fine-grained understanding.The crop is extracted from unlabeled images through an automated data-synthesis pipeline.
4 Experiments
Experiments evaluate Vision-OPD on fine-grained visual benchmarks and separate holdout tasks for generalization, comparing it with agentic, closed-source, and open-source MLLMs. The reported setup uses Qwen3.5-4B/9B models trained on 6.2K synthetic examples.
- Experimental settings: 6.2K synthetic training examples are used to train Vision-OPD on Qwen3.5-4B/9B models with JSD (β = 0.5) as the divergence objective.Top-K distillation with K = 100 reduces the memory overhead of full-vocabulary logit distillation.
- Benchmarks: Fine-grained evaluation covers V* Bench, ZoomBench, HR Bench at 4K and 8K, and MME-RealWorld high-resolution photographs.These benchmarks target tiny targets, zoom-dependent details, high-resolution perception, and real-world visual scenarios.
- Results: Vision-OPD achieves the best overall performance in Table 1, whose entries report accuracy (%) across the compared models.The table highlights the best and second-best results among open-source single-forward-pass models.
- Benchmarks: Holdout evaluation uses MMVP, CV-Bench, MMStar, and POPE to test generalization beyond the training distribution.The holdout group measures general visual understanding and reasoning rather than only fine-grained training-aligned abilities.
- Baselines: Vision-OPD is compared with Thinking-with-Images agents, closed-source models, and open-source models spanning multiple scales.The comparison includes DeepEyes, Thyme, GPT-5.4, Gemini-3.1-Pro, Qwen3.5, and Kimi-K2.6 among others.
4.2 Experimental results
Vision-OPD consistently improves fine-grained visual understanding over alternative training strategies and Qwen3.5 baselines, while preserving performance on holdout tasks. It also surpasses much larger open-source, closed-source, and agentic models.
- Vision-OPD models consistently improve over corresponding Qwen3.5 baselines across all benchmarks.
- Vision-OPD consistently outperforms SFT, RLVR, and OPSD while maintaining strong holdout-task performance.
- Vision-OPD surpasses much larger open-source models and outperforms GPT-5.4 and Gemini-3.1-Pro with the 9B model.
- Despite requiring only a single forward pass, Vision-OPD models outperform “Thinking-with-Images” agentic models.
- Vision-OPD avoids the performance–forgetting tradeoff observed with alternative training approaches.
4.3 Ablation study & analysis
The ablations show that Vision-OPD depends on regularized self-teaching, an effective divergence objective, and sufficiently long on-policy rollouts. Training steadily narrows the regional-to-global gap as the model learns to recover crop-visible evidence from full images.
- Teacher regularization: Without teacher regularization, self-distillation collapses, whereas EMA regularization achieves the highest average of 79.68.The adopted EMA update coefficient is α = 0.05.
- Divergence objective: JSD (β = 0.5) yields stronger gains than forward KL and reverse KL for per-token teacher–student matching.
- Generation length: Increasing generation length from 512 to 1024 tokens improves performance, suggesting longer rollouts provide richer supervision.The study adopts 1024 tokens for subsequent experiments.
- Learning objective: Top-K logits distillation outperforms sampled-token objectives, supporting denser logit-level credit assignment.The selected configuration uses K = 100.
- Regional-to-global gap: Vision-OPD steadily closes the regional-to-global gap during training, indicating improved recovery of crop-visible evidence from full images.Both Vision-OPD models finish with smaller gaps than much larger or closed-source models.
5 Related work
Related work covers on-policy distillation and fine-grained visual understanding in MLLMs. Vision-OPD is positioned against methods that use inference-time cropping and zooming to expose local evidence.
- Fine-grained visual understanding: Figure 5 tracks the regional-to-global gap during VisionOPD training, where a lower gap indicates better recovery of crop-visible evidence from the full image.
- On-policy distillation: On-policy distillation trains students on self-generated trajectories with token-level teacher supervision, reducing train–test mismatch from off-policy distillation.
- Fine-grained visual understanding: “Thinking-with-Images” methods improve fine-grained understanding by gathering visual evidence through inference-time cropping and zooming.
6 Conclusion
Vision-OPD lets an MLLM teach itself from privileged regional inputs, transferring crop-conditioned behavior to full-image rollouts through per-token divergence. Experiments report improved fine-grained understanding without external teachers, labels, verifiers, or inference-time tool use.
- Vision-OPD uses an evidence-centered crop as teacher input and the full image as student input.
- The framework minimizes per-token divergence on student rollouts to provide dense self-supervision.
- Experiments show substantial fine-grained understanding gains over much larger open-source, closed-source, and agentic models.
A Inference speed comparison
Vision-OPD-9B is compared with agentic visual tool-use baselines on inference speed and uses a single forward pass.
- Vision-OPD-9B retains single forward-pass efficiency, unlike the compared agentic and tool-use baselines.
- Vision-OPD-9B achieves the fastest inference speed among DeepEyes, Thyme, DeepEyesV2, and SenseNova-MARS.Speed is measured as the inverse of average per-sample inference time on ZoomBench.
B Prompts
The benchmark prompts combine an image, a question, and answer choices, with response instructions tailored to short-answer or multiple-choice evaluation.
- The complete benchmark prompt set is provided to support reproducibility and future research.
- Multiple-choice prompts require returning only the letter of the selected option, using English or Chinese instructions.
- Several prompts present an image, question, and options, sometimes prefaced by an instruction to select from the choices.
- Short-answer prompts instruct models to respond using a single word or phrase.
C Case Study
A case study shows Vision-OPD-9B correctly reading a small number on a distant boat, while Qwen-3.5-9B fails.
- Vision-OPD-9B answers correctly when reading a small number written on a distant boat.
- Qwen-3.5-9B fails on the same distant-boat reading case, contrasting with Vision-OPD-9B’s result.
- The case study is presented as evidence that Vision-OPD can internalize fine-grained visual understanding capabilities into existing MLLMs.