Contents

On-Policy Visual Evidence Distillation

Shaohang Wei1,‡,*Feifan Song1Guangyue Peng1Wenhao Yu3Wei Li1Wen Luo1Yang Xu4Yufan Shen2Luke Mao2Yang Du2Asher Qin2Houfeng Wang1,†

1Peking University2Tencent3CUHK4Nanjing University

ReVuE overview
Animated SVGPDF

Abstract

Too Long; Don’t Read

    From visual mistakes to better supervision

    Find the first failure. Give the teacher the missing context. Focus learning where that context matters.

    AcquireLook in the right place.
    ReadRead what is actually there.
    GroundUse it to reach the answer.

    Two islands, not one

    Two islands, not one: teacher token scores on the same student trajectory
    Animated SVGPDF

    Better results across two model families

    Weighted-average accuracy (%). Higher is better.

    Hover or focus a value to compare.

    Bold values mark the best OPD result within each model family, including ties. Weighted averages account for benchmark sample counts.

    View the complete results table

    What the trajectories reveal

    The aggregate results raise a practical question: what does a useful reflection actually correct? These cases follow the evidence from the first observation to the final answer.

    Citation (BibTeX)

    arXiv:2609.36838, 2026.

    Open original figure (PDF)