Research map / Vision-language and multimodal / Multimodal LLMs

When Safe Unimodal Inputs Collide: Optimizing Reasoning Chains for Cross-Modal Safety in Multimodal Large Language Models

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces the SSUI dataset of cross-modal implicit reasoning risk examples with reasoning paths and SRPO, a training method that contrasts correct and incorrect reasoning branches via a log-ratio alignment loss plus a reference SFT loss, improving MLLM safety on several benchmarks and a new RSBench.

Problem

MLLMs can be led to harmful outputs when individually safe image and text inputs combine into an unsafe meaning, a failure attributed to difficulty keeping safety alignment through long multi-step reasoning.

Approach

Generates favorable and unfavorable reasoning branches at each step of reference CoT paths, verifies them against ground truth, and optimizes the model with a reference-path likelihood loss plus a step-wise contrastive log-ratio loss over the branch pairs.

Key results

• Qwen2.5-VL-7B + SRPO reaches 76.18 average score on the five safety benchmarks versus 47.58 for the base model, per Table 1.
• LLaVA-NeXT-LLaMA3 + SRPO improves average from 41.12 to 65.99 on the same benchmarks.
• Qwen2.5-SRPO is reported to gain more than 20% absolute in both safety rate and effectiveness rate on RSBench over Qwen2.5-VL.

Surprising findings

• Paper reports SFT underperforms self-explored reasoning methods on harder benchmarks.

Limitations

• Relies on GPT-4o as arbitration judge for RSBench, which may bias evaluation.
• Evaluation limited to two base MLLMs and a focused experimental scope.
• Dataset is small (4,779 samples) and partly manually revised.

Experimental setup

Datasets: SSUI Open Images v7 COCO EgoShots

Benchmarks: USBBench MLLMGuard MSSBench SafeBench VLSBench RSBench

Metrics: ASR ARR Avg ASR PAR ASD SRI safety rate effectiveness rate SER

Baselines: SFT DPO ORPO Claude-3.5-Sonnet Gemini-1.5-Pro Gemini-2.0-Flash GPT-4o Qwen2.5-VL-7B

Builds on: DPO ORPO LoRA SimPO

Closest papers by meaning

• SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization (CVPR 2026 · similarity 0.92)
• Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.92)
• Multimodal Situational Safety (ICLR 2025 · similarity 0.91)
• Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.90)
• Evolving Contextual Safety in Multi-Modal Large Language Models via Inference-Time Self-Reflective Memory (CVPR 2026 · similarity 0.90)
• Pragma-VL: Towards a Pragmatic Arbitration of Safety and Helpfulness in MLLMs (ICLR 2026 · similarity 0.90)