跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2606.25319 · v2

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

首次提交 2026-06-24 10:32
版本更新 2026-11-18 13:59
雷达观测 2026-11-19 19:08
arXiv ↗PDF ↗

摘要

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.

作者备注:revised title and abstract, added more ablation studies, and updated training details and results

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-18 13:59 · 最新观测版本

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

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.

查看此版本解读 →查看此版本原文 ↗