Weilin Wan, Jingtao Han, Weizhong Zhang, Cheng Jin
Accepted at NeurIPS, 2026
This paper establishes the first holistic framework for global MoE architectural optimization, reducing the intractable 16-dimensional search space to a two-phase sequential procedure via mathematical constraints and rank-preserving properties.
Weilin Wan, Fan Yi, Weizhong Zhang, Quan Zhou, Cheng Jin
Published in AAAI, 2026
This paper investigates the interplay between weight pruning and coreset selection, proposing a synergistic training mechanism (SWaST) that alternately optimizes both to improve efficiency and accuracy while preventing critical information loss.
Zhikang Xie, Weilin Wan, Peizhu Gong, Weizhong Zhang, Cheng Jin
Published in AAAI, 2026
This paper proposes Advanced Black-Box Tuning (ABBT), a framework that reduces API calls by an order of magnitude in LLM optimization through instruction-aware initialization and Langevin dynamics-based evolution.
Weilin Wan, Weizhong Zhang, Cheng Jin
Published in NeurIPS, 2025
This paper challenges the assumption that data importance is budget-independent and proposes CADS, a computation-aware data selection framework that dynamically optimizes training subsets under explicit computational constraints.
Weilin Wan, Weizhong Zhang, Cheng Jin
Published in ICME, 2023
This paper presents a skeleton-based video anomaly detection framework (PoMo) that combines memory-augmented pose reconstruction and conditional variational motion prediction to effectively identify spatial and temporal anomalies.
Weilin Wan, Weizhong Zhang, Quan Zhou, Fan Yi, Cheng Jin
arXiv preprint arXiv:2402.18162, 2024
This paper introduces Neural Activation Prior (NAP), a parameter-free post-hoc framework that effectively detects out-of-distribution samples by exploiting the distinct “signal-to-noise” activation patterns within feature channels without requiring retraining.