I’m Weihan Fei, a sophomore undergraduate student at the University of Science and Technology of China (USTC), School of the Gifted Young, majoring in Artificial Intelligence (GPA: 3.87/4.3, Rank: 14/103).

My research interests lie in interpretable and reliable AI systems, especially LLM agents, long-term memory, retrieval/reasoning mechanisms, and human-centered recommendation. I am particularly interested in methods with clear mathematical structure and empirical grounding, where model behavior can be analyzed, verified, and connected to human needs.

I started systematic AI study in July 2025 and have completed full assignment tracks for Berkeley CS61B, Stanford CS229, CS230, and CS224n, with implementations documented in GitHub repositories. I also studied MIT 6.S184 (Generative AI Foundations) and have been reading broadly across LLM reasoning, recommendation, and agent memory.

Research

SHELF: From Similarity Retrieval to Path-Aware Auditable Memory for LLM Agents
Research Intern, USTC, advised by Prof. Xiang Wang
Under Review at NeurIPS 2026

Through a broad survey of LLM-agent memory literature, I identified a coupled bottleneck in existing systems: structure is often used only after similarity retrieval, while reflection gives verbal failure feedback without specifying which access decision should be repaired. I contributed to the theoretical formulation of SHELF, a path-aware auditable memory framework that writes facts into explicit structural addresses and retrieves evidence through query-conditioned paths over entity, facet, time, relation, provenance, and coverage variables. SHELF was evaluated on LoCoMo with Qwen3-8B, Qwen3-32B, and DeepSeek-V3.2, achieving the best Overall F1, BLEU, and LLM-judge scores among compared memory baselines.

Adaptive-Thinking for Generative Recommendation
Research Intern, Alpha-Lab, USTC, advised by Prof. An Zhang

I am investigating the behavioral gap between “think” and “not-think” inference in generative recommendation, motivated by the observation that explicit reasoning can hurt performance on simpler reasoning tasks. I am studying whether these two inference modes rely on different recommendation signals, such as collaborative filtering patterns, popularity bias, user-item affinity, and semantic item descriptions, and exploring criteria for deciding when reasoning should be invoked.

News

• [May 2026] SHELF is under review at NeurIPS 2026.
• [Mar 2026] I started research on SHELF: Path-Aware Auditable Memory for LLM Agents, supervised by Prof. Xiang Wang.
• [Nov 2025] I joined Alpha-Lab at USTC, working on Adaptive-Thinking for Generative Recommendation under the supervision of Prof. An Zhang.
• [Oct 2025] I won the First Prize (Provincial Level, Top 20) at the 17th Chinese Mathematics Competitions, Non-Math Major Track.
• [Sept 2025] I was awarded the Silver Prize of the Outstanding Undergraduates Scholarship