Blogs
Accessible write-ups of my research papers.
The Sim-to-Real Gap of Foundation Model Agents: A Unified MDP Perspective
Why agents that shine on benchmarks stumble in deployment: a unified MDP view of four channels behind the sim-to-real gap, and the remedies for each.
Diagnosing Multi-step Reasoning Failures in Black-box LLMs via Stepwise Confidence Attribution
Stepwise Confidence Attribution (SCA) assigns confidence to each reasoning step from text alone, flagging the steps where a black-box LLM likely went wrong.
UQ & Confidence Calibration in LLMs: A Survey
A taxonomy of uncertainty quantification and confidence calibration methods for LLMs, organized by computational efficiency and four sources of uncertainty.