The program uses hallucination as a visible signal of internal organization, pathway imbalance, temporal reasoning dynamics, memory-factuality conflict, and process-control failure.
Research program
Cognitive Neuroscience of AI Agents
AI agents combine reasoning, memory, tools, interaction, and long-horizon action. My long-term program studies how their intelligence, knowledge, value, and action structures form, organize, evolve, and influence behavior. The near-term focus is how local anomalies propagate across goal, evidence, memory, tool feedback, and action, and how they can be diagnosed and recovered during execution.
Why This Program
A research framework becomes necessary when a new object appears, existing frameworks explain only partial slices, stable problem groups repeatedly arise, a new method combination is required, and the object carries long-term social importance. AI agents now meet these conditions.
A New Object
Agent systems combine perception, language, reasoning, memory, tool use, interaction, long-horizon goals, multi-agent coordination, and emerging self-improvement loops.
Partial Existing Frames
Evaluation maps behavior; interpretability locates mechanisms; safety studies constraints. This program connects behavior, mechanism, knowledge, value, and action.
Stable Problem Groups
Recurring failures expose shared questions about how agent capabilities, internal knowledge, preferences, and actions form, update, and break down.
A Method Combination
Cognitive task design, behavior analysis, representation analysis, mechanism intervention, and process-level control form the methodological core.
Long-Term Importance
Autonomous, social, embodied, and self-improving agents require scientific tools for inspection, guidance, and human-agent coexistence.
The Object: AI Agents as Complex Intelligent Systems
Earlier artificial systems usually had clearer task boundaries, ability boundaries, and behavioral boundaries. Modern AI agents increasingly form abilities through large-scale training, post-training, tool interaction, memory, and environmental feedback. They can transfer across tasks, organize long-horizon goals, call external tools, collaborate with other agents, and influence real environments.
This object calls for a system-level account of internal organization and external action. The research target is the structure of agency itself: how an artificial system organizes intelligence, stores and uses knowledge, develops goal-directed tendencies, and turns model computation into sustained action.
Four Problem Groups
The cognitive neuroscience of AI agents studies four linked structures. Each structure defines a set of empirical questions, mechanisms to locate, and interventions to test.
Intelligence Structure
Reasoning, planning, memory, abstraction, tool use, and long-horizon task organization emerge under training and interaction. I study how these abilities form, generalize, fluctuate, and fail.
- Which internal structures support reasoning, planning, and tool use?
- Which conditions make a capability stable, brittle, or misleading?
- How do post-training and environment interaction reshape capability use?
Knowledge Structure
Agents may store and recombine knowledge in forms that surface prompts reveal only partially. I study where knowledge is organized, how it is called, and how it can be translated into human-checkable evidence.
- Where do factual, procedural, and task-specific structures live internally?
- How do external evidence, parametric knowledge, and personalized memory interact?
- How can internal knowledge be tested and translated into inspectable forms?
Value Structure
Goals, preferences, and behavioral tendencies are shaped by training objectives, reward models, human feedback, deployment context, and social interaction. I treat value as an evolving structure that can be measured and guided.
- How do reward signals and feedback shape stable agent tendencies?
- How do value-related representations interact with task and knowledge structures?
- How can process-level monitoring detect drift in goal-directed behavior?
Action Structure
Tool use, external memory, multi-agent collaboration, embodiment, and environment feedback turn internal computation into action. I study how agent action is organized, monitored, corrected, and updated over time.
- How does model ability become sustained action in external environments?
- How do tool interfaces and memory systems reshape agent behavior?
- How can agents monitor, reflect, correct, and learn from execution failures?
Methodological Loop
Cognitive neuroscience contributes a research workflow for complex intelligent systems. I adapt this workflow to AI agents as a loop from task design to behavioral evidence, internal representation, mechanism intervention, and process-level control.
Task Paradigms
Design tasks that expose capability boundaries, strategy changes, and failure structures.
Behavior Boundaries
Measure when abilities remain stable, when they transfer, and when they collapse.
Representation Analysis
Locate functional parcels, pathway imbalance, temporal dynamics, and memory-factuality conflict.
Intervention Validation
Use ablation, steering, pathway scoring, reward shaping, and environment changes to test mechanisms.
Process Control
Turn diagnosis into monitoring, reflection, corrective action, process reward, and experience updating.
The working loop is: use tasks to define ability, behavior to characterize boundaries, representation analysis to locate structure, intervention to validate mechanism, and process monitoring to connect mechanism with reliable action.
Current Thesis Program
My thesis work uses hallucination as an entry point into this program. Hallucination exposes how internal organization, module mechanisms, knowledge use, memory, reasoning, and long-horizon control can break down. The current papers form a three-level research path: explain the internal coordinate system, locate failures in core agent modules, and regulate long-horizon execution through metacognitive control.
Explain
NeuroCogMap builds an internal coordinate system for model cognition, capability hierarchy, and pathology localization.
Locate
ReDeEP, RHD, and FPPS locate failures in evidence use, temporal reasoning dynamics, and personalized-memory drift.
Regulate
ReARTeR and DS-MCM develop process reward, hierarchical monitoring, reflection, corrective action, and experience updating.
Each project defines a task that exposes a failure, measures behavior, locates internal structure, validates mechanisms, and uses diagnosis to design control.
The current work spans several core agent modules and connects local mechanism diagnosis with long-horizon regulation.
The program provides an explanatory language for where failures form, how they propagate, and how control mechanisms can be designed.
Explain: Internal Cognitive Organization
NeuroCogMap Reveals Cognitive Organization of Large Language Models
arXiv 2026 | Lead project
NeuroCogMap provides the explanatory coordinate system for my thesis work. It treats large language models as experimentally accessible artificial cognitive systems and asks how internal functional organization can explain normal behavior, pathological model behavior, human cortical correspondence, and cognitive-model discovery.
Locate: Agent Module Mechanisms
ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability
ICLR 2025 Spotlight | First author
ReDeEP studies tool and evidence-use hallucination in retrieval-augmented generation. It separates evidence-sensitive pathways from parametric-knowledge pathways, locates failure signals in model internals, and uses this structure for hallucination detection and intervention.
Mechanistic Detection and Mitigation of Hallucination in Large Reasoning Models
ICLR 2026 | First author
RHD moves hallucination analysis from final answers to reasoning trajectories. It characterizes early fluctuation, shallow pattern matching, erroneous backtracking, and pseudo-verification in multi-step reasoning, then uses process-level reward constraints to mitigate these failures.
When Personalization Misleads: Understanding and Mitigating Hallucinations in Personalized LLMs
ACL 2026 Findings | First author
FPPS studies personalization-induced hallucination. It asks how long-term user history and preference signals become entangled with factual representations, and how inference-time representation steering can preserve personalization while reducing factual drift.
Regulate: Metacognitive Control
ReARTeR: Retrieval-Augmented Reasoning with Trustworthy Process Rewarding
SIGIR 2025 | First author
ReARTeR makes retrieval-augmented reasoning reliable at the process level. It combines scalar process rewards, aligned natural-language explanations, bias-aware look-ahead search, and step-level post-training to improve intermediate reasoning and final answers.
Deep Search with Hierarchical Meta-Cognitive Monitoring Inspired by Cognitive Neuroscience
SIGIR 2026 | First author
DS-MCM brings meta-cognitive monitoring into long-horizon search agents. It adds fast monitoring, slower reflective diagnosis, and experience updating so that agents can detect abnormal states, attribute failures, correct themselves, and reuse lessons across tasks.
Trajectory
The program develops from current hallucination mechanisms toward broader scientific tools for increasingly autonomous AI agents. The trajectory moves from systematic failure analysis, to internal knowledge translation, to the explanatory basis for long-term human-agent coexistence.
Understand Systematic Failures
Trace recurring failures from observable behavior back to internal organization, module mechanisms, and process-control breakdowns.
Locate and Translate Internal Knowledge
Study where complex knowledge is organized, how it is called and recombined, how reliable it is, and how it can become human-checkable evidence.
Support Human-Agent Coexistence
Build tools for understanding and guiding agents that form goals, use tools, interact with social systems, and connect action with human values over time.
Future Plan: From Agent Science to AI4Science
The next stage develops this program in two connected directions. First, I will extend the current thesis from hallucination mechanisms to systematic accounts of intelligence, knowledge, value, and action structure. Second, I will use the same experimental loop to study scientific discovery agents. These agents search literature, operate analytical tools and simulators, reason across scales, generate hypotheses, and design experiments. The aim is to make their discovery trajectories inspectable, causally testable, and correctable.
Intelligence Structure
Map how reasoning, planning, memory, abstraction, and tool use form stable capabilities and failure patterns across tasks.
Knowledge Structure
Locate, test, and translate internal knowledge structures so humans can inspect knowledge that surface prompting only partially reveals.
Value and Action Structure
Study how goals, preferences, tool interfaces, social feedback, and environment interaction shape sustained action, then build monitoring and control mechanisms for reliable human-agent coexistence.
Scientific Discovery Agents
Apply the framework to agents that integrate literature search, data analysis, simulation feedback, hypothesis generation, and experiment planning into inspectable discovery trajectories.
AI4Science research loop Scientific task paradigms Discovery behavior and failure boundaries Knowledge and reasoning representation analysis Mechanism intervention with tools or simulators Process monitoring and human-checkable discovery paths