Fine-grained visual reasoning requires multimodal large language models (MLLMs) to identify task-relevant visual evidence and ground their reasoning in local image regions. Existing methods typically rely on reinforcement learning with verifiable rewards or supervised fine-tuning on large-scale annotated reasoning traces, resulting in costly exploration, hand-designed verification rules, or heavy dependence on textual supervision. While On-Policy Distillation (OPD) offers a promising alternative via dense teacher supervision, its token-level objective inherently focuses on local token corrections, failing to assess global trajectory quality. In this work, we first reinterpret OPD as a stop-gradient alignment objective, revealing that its potential is limited by the absence of trajectory-level discrimination. Motivated by this observation, we propose V-Zero, which gates on-policy distillation with contrastive visual evidence. Specifically, V-Zero evaluates each trajectory under paired positive and negative visual views; the resulting likelihood contrast forms a trajectory-level gate that selectively amplifies updates from visually well-supported reasoning paths while suppressing ungrounded ones. Experiments on multiple visual reasoning and perception benchmarks show that V-Zero improves the Qwen3.5-4B backbone by 3.9 points on average while preserving strong generalization. Notably, V-Zero achieves more than 3$\times$ and 6$\times$ training speedups over representative supervised fine-tuning and reinforcement learning baselines, respectively. It requires neither annotated textual reasoning traces nor external reward signals. Code and dataset are released at https://github.com/eVI-group-SCU/V-Zero.
翻译:暂无翻译