← 研踪 PaperTrailarXiv:2611.19786 · v1
Fine-Grained RLAIF for Visual Safety: Learning to Zoom into Hazards for Multimodal Safety Alignment
首次提交 2026-11-15 12:23
版本更新 2026-11-15 12:23
雷达观测 2026-11-17 12:06
摘要
Recent multimodal safety alignment heavily relies on a text-centric paradigm, assuming risks are explicit in language or globally prominent in images. However, real-world hazards are often visually subtle or peripheral, causing existing models to fail due to their reliance on coarse, global visual representations. In this work, we study fine-grained visual safety alignment and introduce a challenging 1,200-sample benchmark across four domains, where hazards necessitate localized inspection. To address this, we propose Fine-Grained RLAIF (FGR), a reinforcement learning framework that casts image zoom-in as a learnable perception policy. FGR employs a decomposed reward design to jointly align three core safety behaviors: (i) where to look via a visual evidence reward, (ii) when to act via a tool-formatting constraint, and (iii) how to reason via visually grounded safety explanations. FGR achieves accuracy ranging from 86.71% to 93.00%, consistently outperforming strong multimodal and agent-based baselines.
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。
v1 · 2026-11-15 12:23 · 最新观测版本
Fine-Grained RLAIF for Visual Safety: Learning to Zoom into Hazards for Multimodal Safety Alignment
Recent multimodal safety alignment heavily relies on a text-centric paradigm, assuming risks are explicit in language or globally prominent in images. However, real-world hazards are often visually subtle or peripheral, causing existing models to fail due to their reliance on coarse, global visual representations. In this work, we study fine-grained visual safety alignment and introduce a challenging 1,200-sample benchmark across four domains, where hazards necessitate localized inspection. To address this, we propose Fine-Grained RLAIF (FGR), a reinforcement learning framework that casts image zoom-in as a learnable perception policy. FGR employs a decomposed reward design to jointly align three core safety behaviors: (i) where to look via a visual evidence reward, (ii) when to act via a tool-formatting constraint, and (iii) how to reason via visually grounded safety explanations. FGR achieves accuracy ranging from 86.71% to 93.00%, consistently outperforming strong multimodal and agent-based baselines.
查看此版本解读 →查看此版本原文 ↗