Projects
Exploring practical AI solutions that bridge cutting-edge research and real-world applications.
MedBench v5
A Dynamic, Process-Oriented, and Hallucination-Aware Benchmark for Clinical Multimodal Models
Project Overview
MedBench v5 is a Chinese medical AI benchmarking platform developed by Shanghai AI Laboratory to systematically evaluate the clinical capabilities and safety of large language models. It extends conventional answer-based evaluation toward dynamic clinical reasoning, process-level assessment, and hallucination awareness.
My Contributions
Led the design of the dynamic clinical reasoning evaluation module, developing three controllable information-flow perturbations (missing information, conflicting evidence, and delayed evidence), a five-stage reasoning audit framework, and hallucination propagation monitoring for multi-turn clinical decision-making.