Research & engineering

About Me

I am a Computer Science undergraduate at Sichuan University. My research interests lie in embodied AI, world models, and intelligent agents, with a particular focus on how autonomous systems perceive, reason, and plan in complex environments.

I am interested in developing reliable and generalizable AI systems by combining structured reasoning, predictive modeling, and learning-based decision-making. I enjoy exploring fundamental research questions through both theoretical understanding and practical experimentation, and I am keen to pursue long-term research collaborations in these areas.

Research Interests

Embodied AI & Robotics

Autonomous systems, planning and decision-making.

World Models & Model-Based Planning

State representations, predictive modeling and planning.

LLM Agents & Reliability

Feedback-driven agents, memory and robust execution.

Multi-Agent Systems

Communication, coordination and structured reasoning.

Selected Projects / Research

Selected engineering work and independent studies. Methods, results and limitations are documented in the linked artifacts.

UAV Inspection Planner

Robotics · Path Planning · Engineering

Contributed topology-aware mission planning and start/end replanning to a team’s powerline-inspection prototype, integrating inspection workflows, visualization and tests. The system operates at mission-planning and simulation level.

GitHub

EgoWM Planning Audit

World Models · Planning · Evaluation

Independently audited action-ranking sensitivity and coupled sampling in an existing action-conditioned world model. The variance-reduction target was not met; navigation success was not evaluated.

GitHub

Dynamic Web-Agent Reliability

LLM Agents · Feedback · Reliability

Studied observation and action reliability in one controlled browser UI. Fresh observations prevented the tested stale actions, while later retries could undo an achieved goal.

GitHub

AgentHijack Timing Extension

Computer Use · Agent Safety · Evaluation

Extended the upstream AgentHijack evaluation with within-step corruption timing controls. A descriptive study of three Chrome tasks examined how stale observations affect execution, with uncontrolled sampling limiting inference.

GitHub Research Note

CPC Memory Mechanism Diagnostic

Agent Memory · State Representation · Reasoning

Used symbolic controls across ten archived worlds to distinguish evidence retention, reader behavior and observational ambiguity. This exploratory mechanism study is separate from the ExpeL/WebShop audit.

GitHub

GTD Execution-Semantics Audit

Multi-Agent Systems · Structured Communication

Independently reproduced the released GTD pipeline and audited communication-graph execution on the fork’s reproduction branch. Coordinate alignment improved graph fidelity without establishing downstream task-performance gains.

GitHub Research Note

ExpeL Memory: Support-Inheritance Audit

Independent extension · Agent Memory · Evaluation

Audited whether inherited support can change solver-visible rule ranking without changing rule text. In one identical 36-rule state, resetting edited-rule support moved its rank from 2 to 15. Six independent compression builds provided a stochasticity control (15 dependent lexical comparisons). Across 50 paired WebShop tasks (100 episodes), successes were 0/50 with inherited support versus 1/50 after reset (exact paired p = 1.0); 96/100 episodes reached the cap. No reliable downstream benefit or equivalence was established.

GitHub

World-Model Planning: Branch Isolation

Independent research · World Models · Planning

Investigated whether executable world-model rollouts preserve parent state while returning the same values as full-copy execution. Implemented WCTR, a restricted Python runtime combining mutation witnesses, alias-aware selective copying and transactional rollback. It recorded zero parent/output contract failures across 33,692 checked transitions under the supported operations. Controlled results were mixed, and native completion tied across four isolation methods. These checks establish neither a general-purpose world-model learning algorithm nor improved downstream planning performance.

GitHub

More research & code on GitHub

Publications

Hyperbolic Self-Supervised Pretraining for Hierarchical EEG Representation Learning in Generalizable Parkinson’s Disease Detection

Jinghong Tang, Zhijing Wu, Yikun Liu, Wei Ju, Jiyao Wang

2026 IEEE International Conference on Bioinformatics and Biomedicine (IEEE BIBM 2026)

Regular Paper, Accepted

Education & Academic Background

Sichuan University

B.Eng. in Computer Science

–

Required-course GPA
4.0/4.0
Average
94.7/100
Rank
2/288

Awards & Honors

  • Oct 2026
    Contemporary Undergraduate Mathematical Contest in Modeling (CUMCM)First Prize, Sichuan Division
  • May 2026
    MCMMeritorious Winner
Technical skills
Programming
Python, C/C++
Machine Learning
PyTorch, PyTorch Geometric, Transformers, Graph Neural Networks
LLM / Agents
SFT, GRPO/GSPO, Tool-Using Agents, Multi-Agent Systems
ML Systems / Tools
Distributed-training concepts (including FSDP), vLLM, Linux, Git, Docker, Conda

News

  1. Received First Prize (Sichuan Division) in the Contemporary Undergraduate Mathematical Contest in Modeling (CUMCM).

  2. Our paper was accepted as a regular paper at IEEE BIBM 2026.

  3. Received the Meritorious Winner (M Award) distinction in the Mathematical Contest in Modeling (MCM).

Contact

I welcome research discussions and opportunities for long-term collaboration.