Xiangchen Song

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

My research studies large language models and sequence models through the lens of identifiable internal representations. The central goal is to make latent representations provably transparent, so that model behavior can be interpreted, attributed, and steered with principled guarantees.

I pursue this goal along the following connected directions:

• Mechanistic interpretability of LLMs: developing identifiable sparse autoencoders (SAEs) with feature consistency for LLM activations, enabling reproducible analysis of internal features and reasoning mechanisms. (blog posts)

• Causal representation learning for nonstationary data: recovering identifiable latent structure and causal relations from time series, video, and text, including both time-delayed and instantaneous dependencies in latent representations. (blog posts)

More recently, I have also been exploring reliable and adaptive agentic LLM systems, including agent evaluation, long-horizon workflows, inference-time control, continual learning and test-time training. This line extends my interest in interpretable and controllable models to LLM systems that reason, act, and adapt over time.

contact

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