MTMCS-Bench: Evaluating Contextual Safety of Multimodal Large Language Models in Multi-Turn Dialogues

TL;DR

MTMCS-Bench tackles the challenge of evaluating contextual safety in multimodal, multi-turn language–vision interactions by introducing two risk setups (escalation-based Type A and context-switch Type B) and a paired safe/unsafe dialogue design across multimodal and unimodal formats. The benchmark comprises 12,032 dialogues and 18,048 QA pairs (30,080 samples) spanning 10 safety scenarios, with MCQ, TF, and open-generation metrics to assess intent recognition, safety-awareness, and helpfulness. Experimental results across 15 open-source and proprietary MLLMs reveal persistent safety–utility trade-offs, with models often failing to detect gradually emerging risk or over-refusing benign queries, and guardrails offering only partial improvements. The work highlights the need for defenses and modeling approaches that explicitly track evolving, context-dependent intent in multimodal dialogue to better balance safety and usefulness in real-world applications.

Abstract

Multimodal large language models (MLLMs) are increasingly deployed as assistants that interact through text and images, making it crucial to evaluate contextual safety when risk depends on both the visual scene and the evolving dialogue. Existing contextual safety benchmarks are mostly single-turn and often miss how malicious intent can emerge gradually or how the same scene can support both benign and exploitative goals. We introduce the Multi-Turn Multimodal Contextual Safety Benchmark (MTMCS-Bench), a benchmark of realistic images and multi-turn conversations that evaluates contextual safety in MLLMs under two complementary settings, escalation-based risk and context-switch risk. MTMCS-Bench offers paired safe and unsafe dialogues with structured evaluation. It contains over 30 thousand multimodal (image+text) and unimodal (text-only) samples, with metrics that separately measure contextual intent recognition, safety-awareness on unsafe cases, and helpfulness on benign ones. Across eight open-source and seven proprietary MLLMs, we observe persistent trade-offs between contextual safety and utility, with models tending to either miss gradual risks or over-refuse benign dialogues. Finally, we evaluate five current guardrails and find that they mitigate some failures but do not fully resolve multi-turn contextual risks.

Figures (28)

• Figure 1: Illustration of two unique setups for assessing MLLM's contextual safety in multi-turn conversations in MTMCS-Bench.
• Figure 2: Demonstration of our three-stage multi-agent data construction workflow (left) and two unique risk setups (right) to assess MLLM multi-turn contextual safety in MTMCS-Bench.
• Figure 3: MCQ, TF, Safety-Awareness (SA), and Helpfulness scores of LLaVA-7B-Instruct, LLaMA-3.2-90B-Vision, and Claude Opus 4.5 under multimodal and unimodal settings. The $x$-axis shows the risk setup (Type A vs. Type B), and the $y$-axis shows the corresponding metric value.
• Figure 4: Prompting strategy for Safety-Awareness score evaluation in multimodal Type A scenarios.
• Figure 5: Prompting strategy for Safety-Awareness score evaluation in multimodal Type A scenarios (Part 2).