Publications

• From Spectra to Joint Schedules in LLM Pre-training: 3+3(+2) Scaling-Law Regimes [arXiv] [code]

  Yichen Wang, Fanghui Liu, Yudong Chen

  [TLDR: We characterize when spectral structure produces power-law learning curves and how learning-rate and batch-size schedules change them, with empirical validation in LLM pre-training.]

  arXiv preprint, 2026
• The φ Curve: The Shape of Generalization through the Lens of Norm-based Capacity Control [arXiv] [code]

  Yichen Wang, Yudong Chen, Lorenzo Rosasco, Fanghui Liu

  [TLDR: We precisely characterize how the generalization learning curve behaves under a suitable model capacity and rigorously prove (Fig. 8.12, CS229 Lecture Notes).]

  39th Conference on Neural Information Processing Systems (NeurIPS), 2025
• Enhancing Exploratory Capability in Visual Navigation using Uncertainty of Implicit Scene Representations [arXiv] [code]

  Yichen Wang, Qiming Liu, Zhe Liu, Hesheng Wang

  37th IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024 (Oral)