TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots
ICML 2026 submission on evaluating clinical fidelity and safety in therapy chatbots.
fangrui-huang
Professional Summary
Sheng Wang is a PhD graduate from The University of Hong Kong, supervised by Prof. Chuan Wu and Prof. Lingpeng Kong.
His research focuses on Agent, LLM Super-Alignment, and Data Synthesis. He has published 14+ papers in top-tier
conferences including NIPS2025 (Spotlight), ICLR2025, ACL2024/2025, EMNLP2025.
Education
PhD Computer Science
2022-09
2025-11
The University of Hong Kong
B.Eng. AI & Automation
2017-09
2021-07
Huazhong University of Science and Technology
Interests
Agent
LLM Super-Alignment
Data Synthesis
Featured Publications
•
•
1 min read
ToolSelf: Runtime Self-Reconfiguration for Robust Long-Horizon Agentic Tasks
ICML 2026 submission on runtime self-reconfiguration for agentic systems (Co-Corresponding Author).
jingqi-zhou
MoS: Unleashing parameter efficiency of low-rank adaptation with mixture of shards
A paper at ICLR 2025 presenting MoS, a method for more parameter-efficient LoRA through mixture of shards.
ProReason: Multi-modal proactive reasoning with decoupled eyesight and wisdom
EMNLP 2025 main conference paper on multi-modal proactive reasoning (Co-Corresponding Author).
jingqi-zhou
QSpec: Speculative decoding with complementary quantization schemes
EMNLP 2025 main conference paper on efficient inference through speculative decoding.
juntao-zhao
TreeSynth: Synthesizing Diverse Data from Scratch via Tree-Guided Subspace Partitioning
A spotlight paper at NeurIPS 2025 presenting TreeSynth, a method for synthesizing diverse data through tree-guided subspace partitioning.
Recent Publications
(2026).
TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots.
ICML 2026 Submission.
(2026).
ToolSelf: Runtime Self-Reconfiguration for Robust Long-Horizon Agentic Tasks.
ICML 2026 Submission.
(2025).
Developing and Utilizing a Large-Scale Cantonese Dataset for Multi-Tasking in Large Language Models.
In EMNLP 2025 Findings.
(2025).
MlingConf: A Comprehensive Study of Multilingual Confidence Estimation on Large Language Models.
In ACL 2025 Findings.
(2025).
UAlign: Leveraging Uncertainty Estimations for Factuality Alignment on Large Language Models.
In ACL 2025 Main.