Research

Publications

Persistent Depth Ordering amid Shifting Block-Bypass Responses in Language Model Pretraining

Shengye Tao, Yinzhu Cheng, Haihua Xie · 2026 · arXiv preprint

arXiv:2610.01165 · PDF

Current Directions

I currently focus on four connected areas:

• Learning theory and model behavior
• Quantitative trading and sequential decision-making
• Cognitive augmentation and human–AI collaboration
• AI agents / research harnesses

Ongoing Research

Training-Stage Behavior and Interpretability of Language Models

I study layer-level behavior, intervention responses, and cross-checkpoint stability across stages of pretraining, with an emphasis on reproducible experimental design, causal interventions, and auditable results.

Identification of Heavy-Tailed Stochastic Processes

I am studying random time changes, subordination, heavy-tailed noise, and data-driven identification of non-Gaussian stochastic processes.

Research Harnesses

I explore how AI agents can organize literature review, problem decomposition, experiments, code auditing, result verification, and paper writing into traceable and testable research workflows.

Current Roles

• Research Assistant Intern, BIMSA (Sep 2026 – Present)
• Youth Assistant, Renmin University of China–Westlake University Joint Institute for Future Humanity (Sep 2026 – Present)
• Undergraduate in Mathematics & AI; participating in the BIMSA Mathematics & AI training program

Notes & CV

Research notes are published on this site. See my current CV for experience and projects.