SafeCap: Improving LVLM Safety with Image Captioning Reinforcement Learning

2026-08-11

AI Summary

This work addresses the vulnerability of large vision-language models (LVLMs) to jailbreaking attacks that bypass safety alignment through malicious image inputs. To mitigate this, the authors propose SafeCap, a novel framework that leverages image captions as an intermediate representation: a policy model trained via reinforcement learning first generates safety-aware image descriptions, which then guide a frozen language model to produce aligned and safe responses—eliminating the need for direct rejection-based supervision. Evaluated across five safety benchmarks, SafeCap achieves average improvements of 3.7–19.0 points over existing methods such as SFT, DPO, and SafeGRPO, while simultaneously maintaining or enhancing performance on six vision-centric utility benchmarks, demonstrating its effectiveness in concurrently ensuring safety and preserving multimodal functionality.

Abstract

Large vision-language models (LVLMs) remain vulnerable to jailbreak attacks that exploit visual inputs to bypass safety alignment inherited from their language backbones. We propose SafeCap, a reinforcement-learning framework that aligns LVLMs through learned self-captioning. SafeCap trains a policy model to first generate a safety-relevant image caption and then produce a final answer; the caption is further optimized by whether it enables a frozen LLM to reach a safety-aligned decision. This caption-mediated objective encourages the policy to expose visual cues relevant to safe response generation rather than relying solely on direct refusal supervision. Across five multimodal safety benchmarks and six vision-utility benchmarks, SafeCap substantially improves aggregate safety performance under its intended DirectCap protocol, with gains of 3.7-19.0 points in safety average across four model settings while maintaining comparable or improved vision utility. Under controlled comparisons on matched backbones and data, SafeCap outperforms safety SFT, DPO, and SafeGRPO, demonstrating the effectiveness of caption-mediated reinforcement learning for multimodal safety alignment.

Problem

• LVLM safety
• jailbreak attacks
• multimodal safety alignment
• visual inputs
• safety alignment

Innovation

• reinforcement learning
• image captioning
• multimodal safety
• safety alignment
• LVLM