I study how large language models represent and manipulate knowledge internally, with a focus on mechanistic interpretability and long-context reasoning. Before this I worked on ____ at ____.
I am always happy to talk about research. The fastest way to reach me is email.
Research Interests
- Mechanistic interpretability
- Long-context reasoning
- Efficient inference and serving
- Evaluation of agentic systems
Current Projects
- Attention Sink Atlas active
Cataloguing where and why attention sinks form across model families, and what breaks when you remove them. code
- Long-Horizon Agent Evals active
A harness for evaluating agents on tasks that take hundreds of steps, where per-step accuracy stops being a useful signal.
- KV Cache Compression paused
Exploring what is actually recoverable after aggressive KV cache eviction. Shelved pending better probes.
Selected Publications
- A Representative Paper Title About Interpretability
Bowen Tian, Co Author, Senior Author
NeurIPS 2026 · Spotlight arXivpdfcode - Another Paper, This One About Long Context
Co Author, Bowen Tian, Senior Author
ICLR 2026 arXiv