Yong Liu

Chen Qian, Dongrui Liu, Haochen Wen et al.

Large reasoning models (LRMs) have demonstrated impressive capabilities in complex problem-solving, yet their internal reasoning mechanisms remain poorly understood. In this paper, we investigate the reasoning trajectories of LRMs from an information-theoretic perspective. By tracking how mutual information (MI) between intermediate representations and the correct answer evolves during LRM reasoning, we observe an interesting MI peaks phenomenon: the MI at specific generative steps exhibits a sudden and significant increase during LRM's reasoning process. We theoretically analyze such phenomenon and show that as MI increases, the probability of model's prediction error decreases. Furthermore, these MI peaks often correspond to tokens expressing reflection or transition, such as ``Hmm'', ``Wait'' and ``Therefore,'' which we term as the thinking tokens. We then demonstrate that these thinking tokens are crucial for LRM's reasoning performance, while other tokens has minimal impacts. Building on these analyses, we propose two simple yet effective methods to improve LRM's reasoning performance, by delicately leveraging these thinking tokens. Overall, our work provides novel insights into the reasoning mechanisms of LRMs and offers practical ways to improve their reasoning capabilities. The code is available at https://github.com/ChnQ/MI-Peaks.

Danni Peng, Yuan Wang, Huazhu Fu et al.

Personalized federated learning (PFL) studies effective model personalization to address the data heterogeneity issue among clients in traditional federated learning (FL). Existing PFL approaches mainly generate personalized models by relying solely on the clients' latest updated models while ignoring their previous updates, which may result in suboptimal personalized model learning. To bridge this gap, we propose a novel framework termed pFedSeq, designed for personalizing adapters to fine-tune a foundation model in FL. In pFedSeq, the server maintains and trains a sequential learner, which processes a sequence of past adapter updates from clients and generates calibrations for personalized adapters. To effectively capture the cross-client and cross-step relations hidden in previous updates and generate high-performing personalized adapters, pFedSeq adopts the powerful selective state space model (SSM) as the architecture of sequential learner. Through extensive experiments on four public benchmark datasets, we demonstrate the superiority of pFedSeq over state-of-the-art PFL methods.