Paper Overview
Field: Computer Vision (CV) Authors: Shiao Xie, Siyu Chen, Jianwei Lv, Xiandong Li, Yujin Wang Published: 2026-10-19 arXiv: 2610.26856
Summary
Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervision with dense token-level corrections from a privileged teacher equipped with an evidence-centered visual view.
However, the resulting corrections may conflate visual evidence with non-visual shifts induced by linguistic priors, response style, or the privileged view itself. The authors formalize this phenomenon as privileged-correction contamination: a token-level correction is contaminated when its update direction cannot be consistently attributed to the privileged visual evidence. Directly distilling contaminated corrections may reinforce erroneous patterns.
Proposed Method: Visually Attributable Gating (VAG)
VAG addresses contamination by:
- Using verifier-derived advantages as optimization anchors
- Validating each privileged correction through contrast between evidence-complete and evidence-degraded teacher views
- Accepting only corrections whose direction is supported by visual contrast, thereby suppressing inconsistent residuals while preserving visually grounded guidance
- Paper: <https://arxiv.org/abs/2610.26856>
Results
Experiments on six visual reasoning benchmarks show that VAG consistently outperforms GRPO and OPSD methods, establishing a new state-of-the-art average performance.