V-Zero: Answer-Label-Free On-Policy Distillation with Contrastive Evidence Gating for Fine-Grained Visual Reasoning

2026-06-23

AI Summary

This work addresses the high training cost in fine-grained visual reasoning caused by reliance on annotated answers or reward-based verification. The authors propose a novel answer-label-free policy distillation framework that introduces, for the first time, a contrastive evidence gating mechanism. This mechanism leverages question-relevant image regions and synthetically constructed negative views to generate unsupervised signals, enabling joint optimization of trajectory-level discrimination and token-level distillation. By integrating on-policy distillation, region cropping, negative view construction, and stop-gradient alignment, the method substantially enhances the model’s ability to exploit local evidence and improves generalization across multiple visual reasoning benchmarks. The approach achieves over 5× faster training than supervised fine-tuning and more than 10× speedup compared to reinforcement learning baselines.

Abstract

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 agentic methods typically rely on reinforcement learning with verifiable rewards or supervised fine-tuning on large-scale annotated reasoning traces, leading to costly exploration, hand-designed verification rules, or heavy dependence on textual supervision. A natural way to avoid such external answer labels is to learn from trajectories sampled by the student itself, which points to On-Policy Distillation (OPD). To understand what OPD can and cannot provide for visual reasoning, we revisit it as negative-free stop-gradient alignment. This perspective shows that, although OPD provides effective token-level correction, its ceiling is constrained by the absence of trajectory-level discrimination. Motivated by these observations, we propose V-Zero, an answer-label-free framework for visual reasoning with contrastive evidence gating. V-Zero uses no annotated textual answer labels; instead, during training it pairs a question-relevant regional crop with a negative visual view to evaluate student-sampled trajectories and gate dense token-level distillation. Experiments on multiple visual reasoning benchmarks show that V-Zero consistently improves fine-grained visual reasoning while preserving strong generalization. Notably, V-Zero is more than 5$\times$ faster than previous supervised fine-tuning methods and more than 10$\times$ faster than reinforcement learning baselines. Code and dataset will be released at https://github.com/eVI-group-SCU/V-Zero

Problem

• fine-grained visual reasoning
• answer-label-free learning
• on-policy distillation
• multimodal large language models
• visual evidence grounding

Innovation

• On-Policy Distillation
• Contrastive Evidence Gating
• Fine-Grained Visual Reasoning
• Answer-Label-Free Learning
• Multimodal Large Language Models