📝 Publications
[CCF-A] PUPPET: Neural-Symbolic Standardized Patients for Mental Health
Chen Xu, Yu Ji, Zhenyu Lyu†, Yang Yi, Yizhe Yang, Luyao Ji, Chaoyi Chen, Xianyang Wang, Tian Lan, Zhihua Wang，Juan Wang, Xunde Dong, Fuze Tian, Qunxi Dong, Bin Hu
(Corresponding author/Second student author)
ACL 2026 Main Oral
- Addressed the high cost and scarcity of human standardized patients in counselor training and the lack of interpretable state control in pure LLM baselines such as prompting, CoT, and memory agents; built a controllable, interpretable patient agent for training human counselors
- Proposed the PUPPET/PUPPET-TRAINER neural-symbolic framework with an Observe–Think–Behave architecture; used LLM Valuator/Generator modules for natural dialogue and flexible expression, and explicitly controlled the intervention–rule–state causal chain with expert rules, probabilistic logic, and probabilistic state machines
- Validated the advantage of combining LLM expressiveness with symbolic controllability across CBT, MI, and 12356 clinical training scenarios; achieved 74.00% MI psychological-state accuracy under DeepSeek-V3.2-chat, +24pp over baseline, and accurately distinguished expert, novice, and high-risk intervention trajectories, improving expert feedback quality and novice training confidence
Preprint First, Do No Harm: AI Supervisor Scaffolds Novice Growth in Counselor Education
Chen Xu, Zhenyu Lyu, Tian Lan, Yang Yi, Yu Ji, Luyao Ji, Jian Shen, Zhihua Wang, Leyang Cui, Jieshuo Zhang, Xiaohua Wan, Qunxi Dong, Minqiang Yang, Juan Wang, Xiuling Liu, Bin Hu
(First student author)
- Defined the AI Supervisor feedback task for novice counselor training: from a multi-turn counselor–client dialogue, the model locates risky utterances (where), identifies violated principles (which), and generates pedagogical feedback for improvement (why); addressed scarce real clinical violations and the limitation that AI counselors optimize answers rather than teach
- Built ETHICSCAFF, a 9,915-sample human-in-the-loop dataset, using 15 novice ethical-violation categories to synthesize multi-turn dialogues between AI novices and simulated clients; trained with SFT + GRPO and proposed Novice Growth Reward, which uses gains in a frozen novice model’s localization/classification ability before and after reading feedback as a proxy RL reward
- Qwen3-14B achieved 94.37\% F1 in violation-principle classification, a 144.1% relative gain over the base model; Qwen3-8B achieved 74.24% F1 in problematic-utterance localization; supervisor feedback improved novice responses across six clinical metrics in simulation and significantly raised eight counseling-task confidence metrics in a real student study Project
[JCR Q1, 中科院一区Top] Bridging the Gap Between Data Distribution and Model: Dynamic Data Distribution Optimization for Improving Critique Capabilities of Large Language Models
Chen Xu*, Tian Lan*, Zhenyu Lyu*, Heyan Huang, Minqiang Yang, Qunxi Dong, Jieshuo Zhang, Bin Hu (Co-First author)
Expert Systems With Applications 2026
- Proposes DIDD, a dynamic iterative data distribution optimization framework that detects model’s vulnerable data distributions and builds tailored training datasets, solving the static data-model mismatch in LLM critique.
- Systematically analyzes the impact of data type and text quality distributions on LLM critique ability, deriving heuristic optimal distribution guidelines for single/pair-wise critique tasks.
- 7B-scale LLMs optimized by DIDD outperform state-of-the-art 7B-13B baselines and even GPT-3.5-turbo on four critique benchmarks with superior data efficiency.
[CCF-A] Voicing Your Emotion: Integrating Emotion and Identity in Cross-Modal 3D Facial Animations
Wenfeng Song, Zhenyu Lyu, Xuan Wang, Xia Hou (First student author)
2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR)
- Proposes a cross-modal 3D talking face generation pipeline fusing speech audio, textual prompts and emotion labels.
- Breaks the limitations of traditional methods that merely synthesize lip movements from speech, enabling 3D avatars with customized appearances and rich emotional depth