Xiangchen Song (宋祥辰)

xiangchen-head-2024.jpeg

I am a PhD student in Machine Learning Department at Carnegie Mellon University, advised by Prof. Kun Zhang (CMU-CLeaR Group). Previously, I studied Computer Science at UIUC with Prof. Jiawei Han.

research

Large language models are sequence models, and my work aims to make their internal representations provably identifiable so we can interpret and steer model behavior with principled guarantees. I build on causal representation learning to recover latent structure in temporal data (time series, video, and text), and bring this lens to mechanistic interpretability of LLMs: designing identifiable sparse autoencoders with feature consistency for LLM activations, analyzing internal reasoning mechanisms, and enabling targeted control for more reliable and efficient model behavior.

contact

Email: xiangchs [at] cs [dot] cmu [dot] edu

news

Sep 23, 2025 Two papers about efficient LLM reasoning have been accepted to NeurIPS 2025 Workshop on Efficient Reasoning (ER@NeurIPS’2025)!!
Sep 22, 2025 Two papers on mechanistic interpretability have been accepted to Mechanistic Interpretability Workshop at NeurIPS 2025 (MechInterp@NeurIPS’2025)!!
Sep 18, 2025 One paper “LLM Interpretability with Identifiable Temporal-Instantaneous Representation” has been accepted to The Thirty-ninth Conference on Neural Information Processing Systems (NeurIPS’2025)!!
May 01, 2025 One paper “Reflection-Window Decoding: Text Generation with Selective Refinement” been accepted to The Forty-Second International Conference on Machine Learning (ICML’2025)!!
Jan 22, 2025 One paper “On the Identification of Temporal Causal Representation with Instantaneous Dependence” has been accepted to The Thirteenth International Conference on Learning Representations (ICLR’2025) with oral presentation!!

selected publications

  1. LLM Interpretability with Identifiable Temporal-Instantaneous Representation
    Xiangchen Song*, Jiaqi Sun*, Zijian Li, Yujia Zheng, and Kun Zhang
    In The Thirty-ninth Annual Conference on Neural Information Processing Systems, Dec 2025
  2. Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEs
    In Mechanistic Interpretability Workshop at NeurIPS (Spotlight), Dec 2025
  3. Causal Temporal Representation Learning with Nonstationary Sparse Transition
    In The Thirty-eighth Annual Conference on Neural Information Processing Systems, Dec 2024
  4. Temporally Disentangled Representation Learning under Unknown Nonstationarity
    In Thirty-seventh Conference on Neural Information Processing Systems, Dec 2023