CV

Research Scientist Intern (part-time)

Aether AI

Research on self-evolving agents through autonomous exploration, verification, and persistent memory.

• RSIAgent combines autonomous exploration with iterative memory refinement, using verified experience from each round to guide the next. Project page.
• Built on open-source GLM-5.3 and Kimi-K3, RSIAgent surpasses GPT-6 Astra in the paper’s reported partial-credit scores: 78.98 vs. 72.60 on OSWorld 2.0 and 84.82 vs. 82.26 on Agents’ Last Exam (Near-term).
• Hugging Face Daily Papers #6 (Sep 15, 2026); Headline feature in Synced (机器之心), an AI news outlet.

Graduate Researcher

University of Illinois Chicago

Research on self-evolving agent skills, verification, and causal representation learning.

• CoEvoSkills evolves reusable skills through co-evolutionary verification; accepted to COLM 2026.
• TRACE recovers continuously evolving causal mechanisms; accepted to ICML 2026.
• MOSAIC discovers interpretable modules in scientific time series; accepted to NeurIPS 2026.
• Worked on verifiable rewards and auditable agent environments.

Undergraduate Researcher

Zhejiang University

Research on dexterous robotic manipulation with Prof. Qi Ye.

• Studied goal-driven hand-object interaction and cross-embodiment retargeting.

Ph.D.

University of Illinois Chicago

Computer Science

• Advisor: Prof. Philip S. Yu
• Collaborator: Prof. Kun Zhang, Carnegie Mellon University
• Research: Recursive Self-improvement of AI Agents and Causal Representation Learning

B.Eng.

Zhejiang University

Automation (Control & Robotics)

• Chu Kochen Honors College
• Outstanding Graduate

RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments

Under review at ICLR 2027

Agents autonomously propose tasks, act, verify outcomes, and retain reusable experience in memory to guide broad-then-deep exploration. This RSI loop lets open-source GLM-5.3 and Kimi-K3 surpass GPT-6 Astra in partial-credit scores on OSWorld 2.0 (78.98 vs. 72.60) and Agents’ Last Exam (84.82 vs. 82.26). Hugging Face Daily Papers #6 (Sep 15, 2026); Headline feature in Synced (机器之心), an AI news outlet.

Agentic Commerce World: An Auditable and Verifiable Environment for Vibe Commerce

arXiv preprint

CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification

Conference on Language Modeling (COLM)

TRACE: Trajectory Recovery for Continuous Mechanism Evolution in Causal Representation Learning

International Conference on Machine Learning (ICML)

MOSAIC: Module Discovery via Sparse Additive Identifiable Causal Learning for Scientific Time Series

Conference on Neural Information Processing Systems (NeurIPS); Accepted

Beyond Math and Code: Lightweight Corpus-Grounded Process Rewards for Factual Question Answering

Under review at ICLR 2027

Corpus co-occurrence yields sentence-level rewards that improve all 30 factual-QA settings and beat four RL baselines in 17 of 20 comparisons. Full training is 5.5–10.4× faster, averaged over Qwen3-4B and Qwen3-8B.

Paying for Honesty Without Knowing the Truth: Reputation-Penalty Design for LLM Marketplace Agents

arXiv preprint

Scaling LLM Agent Learning with Data Synthesis: A Comprehensive Survey

OpenReview preprint