ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents

Xuhao Hu^*, Xi Zhang^*, Haiyang Xu^†, Kyle Qiao, Jingyi Yang, Xuanjing Huang, Jing Shao, Ming Yan^†, Jieping Ye

End-to-end computer-use agent for optimal GUI-tool path selection, with interleaved trajectory scaling, tool-bootstrapped GUI RFT, and online agentic RL.

Paper / Project HomePage / Code / Model

Mobile-Agent-v3.5: Multi-platform Fundamental GUI Agents

Xi Zhang, Haiyang Xu, et al.; Xuhao Hu as core contributor

Responsible for GUI + tool capabilities for next-generation CUAs/Claws, OSWorld-MCP/MobileWorld results, verifiable tasks, rollout sandboxes, and GUI critic pipelines.

Paper / Code / X-PLUG

SALAD-Bench: A Hierarchical and Comprehensive Safety Benchmark for Large Language Models

Lijun Li^*, Bowen Dong^*, Ruohui Wang^*, Xuhao Hu^*, Wangmeng Zuo, Dahua Lin, Yu Qiao, Jing Shao^†

Early LLM-based safety judge models. MD-Judge-v1/v2 and related resources have reached 100K+ Hugging Face downloads and 300+ citations.

Paper / Code / MD-Judge-v1 / MD-Judge-v2

Unveiling Visual Leakage in Multimodal Safety

Xuhao Hu^*, Dongrui Liu^*, Hao Li, Xuanjing Huang^†, Jing Shao^†

Studies visual leakage in multimodal safety and releases VLSBench, an evaluation and alignment resource with scalable multimodal safety data construction.

Paper / Project HomePage / Code / Dataset

IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks

Xiaoya Lu^*, Zeren Chen^*, Xuhao Hu^*, Yijin Zhou, Weichen Zhang, Dongrui Liu^†, Lu Sheng^†, Jing Shao^†

Benchmark for interactive safety of VLM-driven embodied agents in realistic household tasks, covering hazards introduced by multi-step perception, planning, and action.

Paper / Code

LLMs Deceive Unintentionally: Emergent Misalignment in Dishonesty from Misaligned Samples to Biased Human-AI Interactions

Xuhao Hu, Peng Wang, Xiaoya Lu, Dongrui Liu, Xuanjing Huang, Jing Shao^†

Systematically studies emergent dishonesty misalignment under high-stakes settings through direct SFT, downstream SFT, preference learning, and multi-turn human-AI interactions.

Paper / Code

SafeWork-R1: Coevolving Safety and Intelligence under the AI-45 Law

Center for Safe & Trustworthy AI, including Xuhao Hu

Multimodal reasoning model and SafeLadder framework for coevolving safety and general capability. I contributed to multimodal reward models for safety and helpfulness used in multi-target RL.

Paper / Project HomePage

DARE: dLLM Alignment and Reinforcement Executor

Jingyi Yang^*, Yuxian Jiang^*, Xuhao Hu^*, Shuang Cheng, Biqing Qi, Jing Shao^†

An efficient RL training framework for diffusion large language models, integrating dLLM-tailored RL algorithms and practical recipes on top of veRL-style infrastructure.

Paper / Code

Taming Masked Diffusion Language Models via Consistency Trajectory Reinforcement Learning with Fewer Decoding Step

Jingyi Yang, Guanxu Chen, Xuhao Hu, Jing Shao^†

Introduces EOS Early Rejection, Ascending Step-Size decoding, and Consistency Trajectory GRPO to improve masked diffusion language models with fewer decoding steps.

Paper / Code

Frontier AI Risk Management Framework in Practice: A Risk Analysis Technical Report

Shanghai Artificial Intelligence Laboratory, including Xuhao Hu as core contributor

Technical report applying the Frontier AI Risk Management Framework and AI-45 Law to analyze frontier-model risks across cyber offense, bio/chemical risk, persuasion, autonomous AI R&D, deception, self-replication, and collusion.

Paper