This research tackles a critical challenge in deploying medical AI: detecting when vision-language models (VLMs) start performing poorly due to data distribution changes in real-world settings. The authors propose a two-stage approach that uses feature distribution shifts and uncertainty measures to flag degradation without needing ground truth labels, enabling timely model updates and maintaining diagnostic accuracy as imaging conditions or patient populations evolve.
Computer Vision · Guanghao Li, Yijun Wang, Yuhao Huang, Ruiqin Xiong, Xiaochun Cao · 6 min read
This paper introduces a novel approach that combines retrieval and reasoning in large language models using a Monte Carlo Tree Search (MCTS)-inspired method to fetch knowledge that aligns with the generated reasoning path. Unlike traditional retrieval-augmented generation that fetches documents before reasoning, the method iteratively retrieves context that supports the evolving chain of thought, improving answer quality and reducing hallucinations.
Machine Learning · Jiaoyang Li, Shaofeng Zeng, Linyuan Deng, Bo An, Amin Azmoodeh, Zihao Li, Hongyang Zhang, Zhen Wang, Zhi-Hong Deng · 5 min read
This research introduces a breakthrough method for detecting valid mathematical reasoning in large language models by analyzing the spectral properties of their attention mechanisms. Instead of relying on external tools or post-hoc verification, the authors show that successful reasoning exhibits distinct patterns in the eigenvalue distribution of attention matrices, enabling early detection of correct solutions during generation.
Machine Learning · Siyuan Liu, Jiahao Xie, Lianghao Xia, Zhiheng Ma, Conghui He, Ruochen Wang, Ziheng Wang, Jiahao Xie, Lianghao Xia, Zhiheng Ma, Conghui He, Ruochen Wang, Ziheng Wang · 5 min read