Undergraduate researcher · Fudan University
Understanding how AI systems reason.
Building agents that can recover and improve.
I study reasoning models, AI agents, and interpretability at FudanNLP Alex Lab, advised by Prof. Yixin Cao. My work connects the analysis of reasoning processes with methods for guiding agent behavior.
I am an undergraduate at Fudan University and an exchange student at UC San Diego for Fall 2026.
Research focus
How do models reach the same answer through different computational paths?
SliceGraphCan model routing signals guide exploration and the choice of a final solution?
RISA01 / Updates
Released Trajectory Lens, a local tool built on TraceGraph for inspecting raw agent logs and comparing runs.
ARM was selected as an Oral paper at EMNLP 2026.
New paper: Disagree to Explore, Agree to Commit, introducing routing-guided test-time scaling for software agents.
Released From Fluency to Accountability, an open claim-lifecycle analysis of ICLR 2026 peer review.
New paper: MoE Routing as Signal for Reasoning Control.
Released TraceGraph and SliceGraph.
Thinking Traps in Long Chain-of-Thought (TAAR) was accepted to ACL 2026 Findings.
Released CoT Folding, a structural analysis of LLM reasoning trajectories.
Released NEX, a label-free method for Chain-of-Thought selection and model ranking.
Released the TAAR preprint on measurable long-Chain-of-Thought traps and adaptive restart.
Released ARM, a training-free approach to generalist LLM agent merging.
02 / Selected work
Three projects on understanding reasoning processes and using them to improve agent decisions.
Method overview
Routing-Guided Test-Time Scaling for Software Agents
Uses a model’s expert-routing signals to guide a software agent’s next action and select a final code patch, without an external judge or model-weight updates.
44.9% → 48.2% macro-average resolved rate on SWE-bench Verified across six gpt-oss conditions, comparing uniform selection with routing arbitration.
Method overview
Diagnosing and improving agent trajectories
Maps agents’ actions and observations into a shared graph to identify productive paths, recurring traps, and opportunities for recovery.
Five benchmark splits. Profiles access, trap exposure, and repair to reveal differences that aggregate success rates conceal.
Method overview
An activation-based atlas of reasoning processes
Groups chain-of-thought traces by their intermediate activation patterns to study how correct solutions diverge and reconnect along the way.
85.5% of 954 problem–model cells contain correct, same-answer runs that belong to multiple process families.
For the complete publication record, including an EMNLP 2026 Oral paper and work in ACL 2026 Findings:
View all publications03 / Beyond research
A personal record of the exchange process, from application and visa preparation to housing, courses, and life at UC San Diego.