Yiran Yao

Zhi-Hong Qi, Da-Wei Zhou, Yiran Yao et al.

In our ever-evolving world, new data exhibits a long-tailed distribution, such as e-commerce platform reviews. This necessitates continuous model learning imbalanced data without forgetting, addressing the challenge of long-tailed class-incremental learning (LTCIL). Existing methods often rely on retraining linear classifiers with former data, which is impractical in real-world settings. In this paper, we harness the potent representation capabilities of pre-trained models and introduce AdaPtive Adapter RouTing (APART) as an exemplar-free solution for LTCIL. To counteract forgetting, we train inserted adapters with frozen pre-trained weights for deeper adaptation and maintain a pool of adapters for selection during sequential model updates. Additionally, we present an auxiliary adapter pool designed for effective generalization, especially on minority classes. Adaptive instance routing across these pools captures crucial correlations, facilitating a comprehensive representation of all classes. Consequently, APART tackles the imbalance problem as well as catastrophic forgetting in a unified framework. Extensive benchmark experiments validate the effectiveness of APART. Code is available at: https://github.com/vita-qzh/APART

Ziyue Wang, Aomufei Yuan, Yongfu Zhu et al.

Reinforcement Learning from Verifiable Rewards (RLVR) has become the dominant approach for improving mathematical reasoning in large language models, yet current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states. Investigating this structure, we find that at the anchor token (the position immediately before the answer marker), correct rollouts converge naturally because they must produce the same answer (cosine similarity ~0.84), yet each retains residual variance from its unique reasoning path. Encouraging full alignment at this point pushes the model to extract a unified "correct decision" representation, reducing sensitivity to which reasoning path was taken. Based on this observation, we propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference. On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales, supported by ablations on loss type, anchor position, layer depth, and loss weight.