Mian Zhang

Hi, I'm a CS PhD in UTD, advised by Zhiyu Zoey Chen. I was a research intern at Tencent AI Lab and a member of the NLP group at Soochow University.

Email  /  CV  /  Twitter  /  Github

Research

My current research focuses on enhancing large language models (LLMs) for social good, with an emphasis on health-related domains. I am actively working on post-training (reinforcement learning and preference learning) and synthetic data generation to equip LLMs with complex and reliable abilities. Besides, I am interested in language agents and vision LLMs.

I am looking for a Research Intern position next summer. Feel free to reach out!

We create a high-quality preference dataset, Psycho-Preference with 36k preference comparison pairs, that align with the preference of professional therapists and use preference learning to train helpful assistants for psycho-counseling.

We present the first systematic benchmark to evaluate LLMs' efficacy for CBT. The key finding is while LLMs excel at recalling CBT knowledge, they struggle with complex tasks requiring deep cognitive analysis and patient-specific responses.

We propose IDEA, a holistic LLM agent framework integrating Induction, DEduction, and Abduction, to generate hypotheses, devise plans, and revise hypotheses iteratively to learn new rules in novel environments and RULEARN, a novel benchmark to assess the rule-learning abilities of LLM agents in interactive settings.

We perform a comprehensive review of LLM-based methods for disease diagnosis.

We propose a comprehensive dataset for studying the inconsistencies in dialogue, which covers the life span of inconsistencies, namely introduction, understanding, and resolution.

We construct a high-quality dataset called SafeConv for conversational safety and train a powerful utterance rewriter to correct unsafe content with Reinforcement Learning from Human Feedback (PPO).

We propose friend-training, a cross-task self-training framework, where models trained to do different tasks are used in an iterative training, pseudo-labeling, and retraining process to help each other for better selection of pseudo-labels.

We propose a pairing enhancement approach for ASTE, which incorporates contrastive learning during the training stage to inject aspect-opinion pairing knowledge into the triplet extraction model.

We propose a mixture of experts network that leverages variable-length context windows when predicting the emotion of different utterances.

Teaching