Overview
- Field: Computer Vision (CV)
- Authors: Shravan Venkatraman, Omkar Thawakar, Ritesh Thawkar
- Paper: arXiv:2508.03807
- Self-improvement for multimodal large language models (MLLMs) is typically driven by reward-based methods that provide only coarse scalar feedback. Distillation offers richer dense token-level supervision, but in the visual domain it usually depends on privileged context built from external annotations, tools, or stronger models.
- The authors introduce CVPD (Contrastive Counterfactual Visual Process Distillation), which, to the best of their knowledge, is the first fully self-contained framework for dense, on-policy, token-level visual self-distillation for MLLMs.
- CVPD identifies visual blind spots: regions where zooming into a region changes and sharpens the model's answer distribution, while removing the same region leaves full-image behavior largely unchanged.
- Such regions expose perceptual information the model can encode but fails to consistently exploit under full-image conditions.
- A three-stage counterfactual criterion identifies these regions directly from the model's own responses and converts them into dense contrastive supervision for self-distillation — no external teacher or annotation required.
- OCRBench: +3.60
- MMStar fine-grained perception: +3.38
- MMStar logical reasoning: +3.08
- Performance is maintained or improved on broader multimodal benchmarks.
Key points
Results
Experiments on Qwen3-VL-8B-Instruct show that CVPD surpasses six self-evolution baselines (including methods relying on external GPT-4o supervision) across 12 benchmarks, with no performance regressions: