ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation

Kanghui Tian^*1,2, Siyuan Liu^*3, Ziang Yan¹, Sheng Xia³, Shuai Dong², Yi Wang^†1
1Shanghai AI Laboratory, ²Fudan Univerisity, ³Nanjing University
^*Equal contribution   ^†Corresponding author

Paper  |  Quick Start  |  Citation

Introduction

This paper introduces ViCuR, a visually grounded privileged-teacher distillation framework for multimodal reasoning. It replaces conventional answer-based privileges with visual cues (query-related evidence in the input) and introduces a lightweight cue recovery module that uses dedicated sink-token cross-attention during prefill to aggregate task-relevant visual evidence into an internal representation, without changing the inference interface or requiring auxiliary losses.

Highlights

• Visual Cue Privilege: Replaces answer/rationale-based teacher privilege with visually grounded cues, reducing the train-test mismatch inherent in conventional OPSD.
• SinkTrack Cue Recovery: A dedicated cross-attention module at selected transformer layers that aggregates task-relevant visual evidence into the sink token during prefill — no extra decoding overhead.
• Consistent Improvements: +1.19 (2B) and +1.24 (8B) over answer-based OPSD, and +0.64 (2B) / +1.08 (8B) over stronger-teacher OPD, across seven multimodal reasoning benchmarks.

Results

Main results with Qwen3-VL-2B and 8B students on in-domain, near-domain, and out-of-domain benchmarks. ViCuR consistently improves the overall average over the corresponding distillation baseline (OPSD or OPD) at both 2B and 8B scales.

Repository Structure

This codebase is built upon verl. Our modifications include:

vicur/
├── modeling_qwen3_vl.py   # SinkTrack cross-attention for Transformers (student training)
└── qwen3_vl.py            # SinkTrack cross-attention for vLLM (student rollout)

examples/on_policy_distillation_trainer/
├── run_qwen3_vl_example.sh   # OPSD/OPD training with ViCuR on Qwen3-VL
└── ...

Key contributions to the verl framework:

• On-Policy Distillation (OPD/OPSD): Training scripts and configuration for on-policy distillation with multimodal support.
• SinkTrack Cross-Attention: Implementation of the cue recovery module in both Transformers and vLLM backends.
• Teacher Prompt Support: Added teacher_prompt dataset field for passing visual cue text to the teacher model during distillation.

Quick Start

1. Environment Setup

Set up the base environment following the verl official installation guide. Specifically, to ensure the correct environment configuration, pull the designated Docker image version by running:

docker pull verlai/verl:vllm011.latest

Then, replace the Qwen3-VL model code in your local Transformers and vLLM installations with the SinkTrack variants provided in vicur/ (including model class registration).

2. Run Training

cd examples/on_policy_distillation_trainer
bash run_qwen3_vl_example.sh

Key configurations in the training script:

• STUDENT_MODEL_PATH: path to the student model (with SinkTrack cross-attention weights initialized)
• TEACHER_MODEL_PATH: path to the teacher model
• distillation.teacher_prompt_key=teacher_prompt: enables visual cue privilege from the dataset

Citation

@article{tian2026vicurvisualcuesrecoverable,
    title={ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation}, 
    author={Tian, Kanghui and Liu, Siyuan and Yan, Ziang and Xia, Sheng and Dong, Shuai and Wang, Yi},
    journal={arXiv preprint arXiv:2606.05718},
    year={2026},
}

License

This project is licensed under the Apache License 2.0, following the license of the original verl framework.

Acknowledgements

This codebase is built upon verl. We thank the authors for their open-source contributions.