Progressive-View On-Policy Distillation for Regional-to-Global Transfer in Multimodal LLMs

Shanfeng Huang; Zhou Fang; Song Xiao; Hai Du

cs.CVSep 26, 2026

Progressive-View On-Policy Distillation for Regional-to-Global Transfer in Multimodal LLMs

Authors: Shanfeng Huang, Zhou Fang, Song Xiao, Hai Du

Organizations: Baidu Inc.

Abstract

Regional-to-global distillation uses crop-conditioned guidance to improve full-image understanding. The challenge is to effectively transfer the teacher's crop-based advantage to the student's full-image inference. We propose progressive-view on-policy distillation (PVD), which shifts the student's view distribution from the crop toward the full image through an intermediate aspect-preserving padded crop. The padded crop preserves regional content while matching the full image's visual-token grid. Across stages, the view mixture assigns increasing probability to the full image. A lightweight regional-advantage weighting reallocates token-level supervision using the crop-conditioned teacher-student log-probability gap. Evaluated under each sampled input, it applies mild reweighting when the gap is small and emphasizes higher-gap tokens when the gap widens. A Jensen-Shannon metric decomposition interprets this schedule as a shift from matched-input imitation toward the deployment objective. Across benchmarks spanning perception, visual mathematics and general multimodal question answering, PVD-full reaches an average accuracy of 77.51 over three seeds, improving on the reward-free distillation baseline by 2.01 points and on its reward-matched variant by 1.00 point. In the reward-free setting, PVD-distill still gains 1.16 points.

Figures & tables

Figure 1: Progressive-view on-policy distillation. Student training uses stagewise mixtures of a regional crop, a grid-aligned padded crop containing the same regional content, and the full image, with sampling mass shifting toward the deployment view while the teacher remains conditioned on the privileged crop. View-conditioned teacher–student matching produces the top- $K$ JSD and regional-advantage (RA) token weights. At deployment, only the student processes the full image.

	Perception & spatial understanding				Visual math		General & diagram QA
Method	ZoomB.	HR-4K	HR-8K	CV-B.	MVista	MVerse	MMStar	MMMU	AI2D	All
Qwen3.5-4B	48.76	83.88	80.62	88.10	78.50	67.61	73.07	66.00	80.73	74.14
GRPO	57.51 $\mathord{\pm}$ .59	84.63 $\mathord{\pm}$ .63	80.25 $\mathord{\pm}$ .38	87.81 $\mathord{\pm}$ .16	76.70 $\mathord{\pm}$ .60	63.10 $\mathord{\pm}$ .19	72.02 $\mathord{\pm}$ .57	63.59 $\mathord{\pm}$ .61	85.11 $\mathord{\pm}$ .31	74.53 $\mathord{\pm}$ .30
V-OPD [ 20 ]	57.99 $\mathord{\pm}$ .71	81.92 $\mathord{\pm}$ .81	78.92 $\mathord{\pm}$ .44	86.47 $\mathord{\pm}$ .23	77.37 $\mathord{\pm}$ .75	68.02 $\mathord{\pm}$ .28	75.13 $\mathord{\pm}$ .70	67.85 $\mathord{\pm}$ .72	85.82 $\mathord{\pm}$ .42	75.50 $\mathord{\pm}$ .31
Static-view distill	56.96 $\mathord{\pm}$ .60	81.83 $\mathord{\pm}$ .69	79.79 $\mathord{\pm}$ .51	85.97 $\mathord{\pm}$ .19	78.20 $\mathord{\pm}$ .80	69.35 $\mathord{\pm}$ .27	75.07 $\mathord{\pm}$ .64	67.78 $\mathord{\pm}$ .78	86.03 $\mathord{\pm}$ .37	75.67 $\mathord{\pm}$ .28
PVD-distill w/o RA	57.71 $\mathord{\pm}$ .65	83.42 $\mathord{\pm}$ .56	80.79 $\mathord{\pm}$ .31	86.57 $\mathord{\pm}$ .18	78.73 $\mathord{\pm}$ .55	70.95 $\mathord{\pm}$ .24	75.33 $\mathord{\pm}$ .67	67.22 $\mathord{\pm}$ .67	86.29 $\mathord{\pm}$ .28	76.34 $\mathord{\pm}$ .29

Table 1: Accuracy (%) on nine benchmarks in three task families. Backbone: Qwen3.5-4B. V-OPD abbreviates Vision-OPD. Post-trained rows report mean

\pm

SD over three seeds and share one training configuration and a

2048

-token evaluation budget. Static-view distill matches the curriculum’s expected view exposure but removes its ordering. Best means are bold, and family averages are reported in the text.

Stage	$\langle g\rangle$ ( $10^{-2}$ )	high-gap	$\langle\ell\rangle$ ( $10^{-3}$ )	weights
1	2.4	5.9%	4.4	$[0.97,1.88]$
2	5.2	12.5%	12.2	$[0.90,1.98]$
3	4.6	12.0%	10.2	$[0.90,1.98]$

Table 2: Stage diagnostics from the reward-free run. We report the mean gap

\langle g\rangle

, the fraction of high-gap tokens with

g_{t}>0.10

, the mean unweighted JSD

\langle\ell\rangle

, and the min–max range of normalized RA weights.

Objective	all correct	all wrong	disagree	zero PG grad.
GRPO	43.1%	23.8%	33.1%	66.9%
V-OPD w/ GRPO	42.4%	13.5%	44.1%	55.9%
PVD-full	41.0%	12.7%	46.3%	53.7%

Table 3: Rollout-group composition averaged over the last

10

training steps. Only disagreeing groups produce a GRPO gradient.

Knowledge Distillation

Knowledge distillation transfers useful behavior or representations from larger or specialized models into more compact or adaptable learners.

See all 821 papers

Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation
May 18, 2026Qianhao Yuan, Jie Lou, Xing Yu +4Multimodal Large Language Models On-Policy Distillation
Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Xiaohongshu Inc.
Token-level Response-visual Attention Guidance for Multimodal LLMs Knowledge Distillation
Jul 1, 2026Jaehyun Jang, Eunseop Yoon, Hee Suk Yoon +3Cross-Modal Knowledge Distillation VLM Distillation
Korea Advanced Institute of Science and Technology, Daejeon, Republic of Korea · University of Illinois Urbana-Champaign, Champaign, USA
Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge
Jun 25, 2026Thomas Shih-Chao Liang, Zhuoran Yu, Yong Jae LeeCross-Modal Knowledge Distillation Fine-Grained Image Classification
University of Wisconsin-Madison

May 18, 2026J/K move · Enter open · S save

Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation

Qianhao Yuan, Jie Lou, Xing Yu, Hongyu Lin+3

Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Xiaohongshu Inc.

Multimodal Large Language Models On-Policy Distillation Multimodal Foundation Models Knowledge Distillation Fine-Grained Visual Perception Multimodal Understanding

Open paper AskPDF arXiv

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

Abstract

Knowledge Distillation

Explore similar work

Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation

Token-level Response-visual Attention Guidance for Multimodal LLMs Knowledge Distillation

Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge