Research map / Multi-agent systems and games / LLM agents
MLA-Trust: Benchmarking Trustworthiness of Multimodal LLM Agents in GUI Environments
Xiao Yang, J.X. Chen, Jun Luo, Zhengwei Fang, Yinpeng Dong, Hang Su, Jun Zhu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| turn single-step MLLM into a multi-step GUI agent | GPT-4o · Safety refusal rate | refusal rate (%): 90.5 -> 70.2 | Hurt |
| turn single-step MLLM into a multi-step GUI agent | Gemini-Pro · Safety refusal rate | refusal rate (%): 86.0 -> 62.5 | Hurt |
| turn single-step MLLM into a multi-step GUI agent | Claude-3-7-sonnet · Safety refusal rate | refusal rate (%): 78.0 -> 57.8 | Hurt |
| let the model execute its first action step (step 0 to step 1) | MLAs · Safety refusal rate | refusal rate (%) | Hurt |
| move from predefined-process tasks to contextual-reasoning tasks | GPT-4o · Safety RtE | RtE (%): 72, 67, 60 -> 39, 52 | Hurt |
| move from predefined-process tasks to contextual-reasoning tasks | LLaVA-NeXT · Safety RtE | RtE (%): 29, 39, 31 -> 30, 38 | No clear effect |
| use proprietary GPT-4o instead of open-source Pixtral-12B | MLA · Truthfulness Interact (T.7) | accuracy (%): 0.0 -> 26.0 | Improved |
| use proprietary GPT-4o instead of open-source Pixtral-12B | MLA · Safety Gen-Jailbreak (S.6) | ASR (%): 42.00 -> 15.00 | Improved |
| ask for speculative email behavior instead of overcompletion email task | Qwen2.5-VL · Controllability | ASR (%): 45.0 -> 57.0 | Hurt |
| increase model scale (LLaVA-OneVision 72B, DeepSeek-VL2 27B, Pixtral-12B vs smaller) | open-source MLAs · Overall trustworthiness ranking | rank | Improved |
Closest papers by meaning
- RiOSWorld: Benchmarking the Risk of Multimodal Computer-Use Agents (NeurIPS 2025 · similarity 0.89)
- OSWorld-MCP: Benchmarking MCP Tool Invocation In Computer-Use Agents (ICLR 2026 · similarity 0.88)
- To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
- ST-WebAgentBench: A Benchmark for Evaluating Safety and Trustworthiness in Web Agents (ICLR 2026 · similarity 0.87)
- Open CaptchaWorld: A Comprehensive Web-based Platform for Testing and Benchmarking Multimodal LLM Agents (NeurIPS 2025 · similarity 0.87)
- OpenAgentSafety: A Comprehensive Framework For Evaluating Real-World AI Agent Safety (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
