Benchmarking LLMs on Authentic Cases from Medical Journals

Wanlong Liu, Junying Chen, Yunjin Yang (Co-first author), Prayag Tiwari, Wenyu Chen, Benyou Wang

Association for Computational Linguistics (ACL) 2026 Accepted

This work introduces ClinBench, a challenging medical benchmark comprising over 2,000 authentic, expert-reviewed clinical cases across 11 specialties. Supporting both textual and multimodal evaluations, comprehensive testing of over 20 LLMs reveals that while human experts excel in specialized fields, LLMs demonstrate superior overall reasoning capabilities in real-world clinical scenarios.

DentalGPT: Incentivizing Complex Multi-modal Diagnosis in Dentistry

Zhenyang Cai, Jiaming Zhang, Junjie Zhao, Ziyi Zeng, Yanchao Li, Jingyi Liang, Junying Chen, Yunjin Yang, Jiajun You, Shuzhi Deng, Tongfei Wang, Wanting Chen, Chunxiu Hao, Ruiqi Xie, Zhenwei Wen, Xiangyi Feng, Zou Ting, Jin Zou Lin, Jianquan Li, Liangyi Chen, Junwen Wang, Shan Jiang, Benyou Wang

Association for Computational Linguistics (ACL) 2026 Accepted

This work presents DentalGPT, a 7B dental-specific multimodal model developed through high-quality domain data and staged training. Using DentalVQA, the largest annotated dental image dataset, together with GRPO-based post-training, the model achieves state-of-the-art results in dental disease classification and VQA, showing the effectiveness of targeted domain adaptation.

[Paper] [Code]

Robust Fault Detection for EV Charging Piles: Integrating Fine-Tuned LLMs with Adaptive Unsupervised Learning

Yunjin Yang, Guibin Wang, Xian Zhang, Jing Qiu, Songjian Chai, Mohamed Abdelkarim Abdelbaky

IEEE Transactions on Consumer Electronics (TCE) 2026 Accepted

This work presents EVCPFD-LLM, a pioneering time-series large language model framework designed to address high-dimensional data and label scarcity via autoencoder-enhanced feature extraction and LoRA-based fine-tuning. Utilizing a dynamic threshold mechanism for fault detection, the model achieves a 13.44% accuracy improvement and 70.87% latency reduction, demonstrating the efficiency of time-series LLMs.

Improving Video Moment Retrieval by Auxiliary Moment-Query Pairs With Hyper-Interaction

Runhao Zeng, Yishen Zhuo, Jialiang Li, Yunjin Yang, Huisi Wu, Qi Chen, Xiping Hu, Victor C. M. Leung

IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 2025 Accepted

This work introduces HyperAux, a hypergraph-based framework that enriches video moment retrieval (VMR) training data by generating and selecting high-quality moment–query pairs. By modeling multi-modal hyper-interactions among frames and queries, HyperAux produces context-aware representations that refine moment boundaries and improve pair selection. Without requiring human annotations, it leverages semantic discrepancies inside and outside moments to guide training. The resulting auxiliary data boosts twelve VMR models across fully-supervised, weakly-supervised, and zero-shot settings on ActivityNet Captions, Charades-STA, and QVHighlights.

[Paper]