Papers by Mingchuan Zhang
SecureWebArena: A Holistic Security Evaluation Benchmark for LVLM-based Web Agents (2026.findings-acl)
Copied to clipboard
Zonghao Ying, Yangguang Shao, Jianle Gan, Gan Xu, Wenxin Zhang, Quanchen Zou, Junzheng Shi, Zhenfei Yin, Mingchuan Zhang, Aishan Liu, Xianglong Liu
| Challenge: | Existing security benchmarks only cover user-level prompts and environmental threats . however, these models are vulnerable to pop-up attacks and prompt injections . |
| Approach: | They propose a security benchmark that covers a set of six attack vectors that span both user-level and environment-level manipulations. |
| Outcome: | The proposed security benchmarks cover a set of six real-world web environments with 2,970 adversarial trajectories and a multi-layered evaluation protocol dissecting agent failures across internal reasoning, behavioral execution, and task outcomes. |
Training Language Model to Critique for Better Refinement (2025.findings-acl)
Copied to clipboard
Tianshu Yu, Chao Xiang, Mingchuan Yang, Pei Ke, Bosi Wen, Cunxiang Wang, Jiale Cheng, Li Zhang, Xinyu Mu, Chuxiong Sun, Minlie Huang
| Challenge: | Large language models (LLMs) have remarkable evaluation and critique capabilities, providing insightful feedback and identifying flaws in various tasks. |
| Approach: | They propose a framework to train critic models using refinement signals to generate feedback loops where critiques guide the model in refining its responses. |
| Outcome: | The proposed framework outperforms traditional methods and open-source models in terms of critique quality and refinement outcomes. |
Adversarial Preference Learning for Robust LLM Alignment (2025.findings-acl)
Copied to clipboard
Yuanfu Wang, Pengyu Wang, Chenyang Xi, Bo Tang, Junyi Zhu, Wenqiang Wei, Chen Chen, Chao Yang, Jingfeng Zhang, Chaochao Lu, Yijun Niu, Keming Mao, Zhiyu Li, Feiyu Xiong, Jie Hu, Mingchuan Yang
| Challenge: | Modern language models rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors, but they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation; (2) the vast diversity of potential adversarials; and (3) the risk of feedback bias and reward hacking. |
| Approach: | They propose an iterative adversarial training method that incorporates three key innovations to address these challenges. |
| Outcome: | Experiments on Mistral-7B-Instruct-v0.3 show that the proposed method significantly enhances robustness and reduces harmful outputs from 5.88% to 0.43%. |