Large-model agents | Hallucination | Mechanistic interpretability | Search

Zhongxiang Sun (孙忠祥)

I am a Ph.D. candidate at the Gaoling School of Artificial Intelligence, Renmin University of China, advised by Prof. Jun Xu. My research focuses on trustworthy LLM agents, especially hallucination mechanisms, reasoning and planning, information retrieval, explainability, and legal AI. I use cognitive-neuroscience-inspired task design, behavior analysis, representation analysis, mechanism intervention, and process-level control to study these systems.

I expect to graduate in June 2027 and am seeking research positions. If you are interested in trustworthy LLMs and agents, their synergy with information retrieval, applied mechanistic interpretability, or cognitive-neuroscience-inspired AI, I would welcome the opportunity to discuss.

Email Google Scholar GitHub
Zhongxiang Sun

Impact Snapshot

Contribution A thesis-level program on trustworthy LLM agents

The work connects internal organization, module-level hallucination mechanisms, and long-horizon metacognitive control into one coherent research program.

Evidence First-author work across ICLR, SIGIR, and ACL

Representative projects cover retrieval-augmented generation, large reasoning models, personalized LLMs, process rewarding, and deep-search agents.

Recognition Spotlight, Best Paper, and Paper Digest highlights

Selected work has been recognized by ICLR, SIGIR-AP, and Paper Digest.

Public Assets NeuroCogMap website, code, and full publication record

The site links project pages, source artifacts, and a compact full publication list for fast faculty-level evaluation of scope, continuity, and impact.

Research

My research studies large-model agents as complex artificial intelligent systems that perceive context, reason over goals, call tools, use memory, interact with humans and other agents, and act in changing environments. I build cognitive-neuroscience-inspired methods for explaining and controlling these systems: task paradigms reveal behavioral boundaries, representation analysis locates internal structure, mechanism intervention validates causal pathways, and process-level monitoring turns diagnosis into control.

Explain

Internal Cognitive Organization

I build NeuroCogMap-style internal cognitive maps that connect sparse features, functional parcels, capability hierarchy, and pathology localization.

Explore internal organization
Locate

Agent Module Mechanisms

I analyze tool evidence use, temporal reasoning dynamics, and personalized-memory drift through ReDeEP, RHD, and FPPS.

View module mechanisms
Regulate

Metacognitive Control

I design process reward, hierarchical monitoring, reflection, corrective action, and experience updating through ReARTeR and DS-MCM.

View metacognitive control

Selected Publications

Representative work with project pages. The complete publication list is maintained separately as a compact text list.

Full Publication List

Education

Awards

Teaching

Experience

Talks