Yiran's Homepage

Senior undergraduate, Fudan University · Mechanistic interpretability & LLM evaluation

prof_pic.jpg

CS, Fudan University

Shanghai, China

I am Yiran Xu (徐一冉), a senior undergraduate in Computer Science at Fudan University.
My work centers on understanding how large language models internally organize computation—
specifically, why modular structures sometimes emerge, when they collapse into shared subspaces,
and how activation geometry encodes reasoning quality.

I am advised by Prof. Yixin Cao (Fudan University) and Prof. Robert Dick (University of Michigan).
Across both groups, I study:

  • the mechanisms behind latent experts, modularity, and low-rank bottlenecks in Transformers
  • the use of activation subspaces to evaluate long-form reasoning, creativity, and process quality
  • how optimization dynamics (e.g., gradient flow, efficiency biases) shape representational structure

I am applying to PhD programs for Fall 2026, aiming to build a principled understanding of
how reasoning is represented inside LLMs and how modularity can be encouraged or reliably induced.

Research interests

  • Mechanistic interpretability of Transformers and LLMs
  • Latent modularity, functional conflict, and emergent experts
  • Representation structure: low-rank manifolds, subspaces, causal directions
  • Evaluation of long-form reasoning and creativity (LongWriter, HelloBench, WritingBench)
  • Causal RL and self-consistent data generation for reasoning tasks

Contact

📧 yiranxu22 [at] m.fudan.edu.cn