Mechanistic interpretability
Localizing sparse model components, causal features, and internal routing paths that explain downstream behavior.
Software Engineering Undergraduate
Xiangkun Sun
I study how modern AI systems make decisions, where those decisions become fragile, and how interventions can turn mechanistic clues into reliable understanding.
Persuasion circuit map
Attention routing, latent choice geometry, intervention
Northeastern University, China
B.Eng. Software Engineering
4.23 / 5.00
Current GPA
3 / 285
Class rank
2023 - Present
Undergraduate study
Research Direction
Mechanisms for trustworthy AI behavior.
LLM persuasion and safety
Studying how persuasive evidence can redirect model choices, and how compact interventions can monitor or block that shift.
Algorithmic problem solving
Building on competitive programming experience to design careful experiments, efficient tooling, and robust evaluation pipelines.
Publication
Latest work on persuasion-induced factual errors.
How LLMs Are Persuaded: A Few Attention Heads, Rerouted
The paper identifies a compact causal mechanism behind persuasion-induced factual errors: a small group of mid-layer attention heads routes answer options through a low-dimensional choice geometry. Persuasion redirects attention toward a target option, and intervention on a rank-one evidence-routing feature can steer or block the effect.
Show compact abstract
The mechanism appears across open-source LLMs and realistic poisoning settings such as Generative Engine Optimization, suggesting persuasion can be framed as a narrow, monitorable circuit rather than a diffuse loss of belief.
Honors
Competition results shaped by algorithms and execution.
ICPC Regional Contest (Shanghai), Silver Medal
CCPC Invitational Contest (Northeast), Gold Medal
RoboCom Developer Competition, National First Prize
Toolkit
Programming, modeling, and communication.
Python
C++
R
Go
Machine Learning
Data Structures
Algorithms
Academic Writing
English
German