Hi, I'm Mingfei. 👋
I'm a PhD student at the University of Technology Sydney, advised by Prof. Yi Zhang and co-supervised by Dr. Mengjia Wu. My research is on Multi-Agent Systems and Agent Memory for large language models — broadly, how should agents remember, reason, and coordinate so that they actually help people?
Always happy to chat about agent memory architectures, multi-agent systems, and legal agentic systems. Reach out →
🔥 What's new
-
Joined MemTensor's Foundation Model Group as an intern.
-
Sep 2026New preprint: PairPref: When Should Memory Guide the Answer? A Benchmark for Contextual Preference Use — evaluating when assistants should apply or set aside remembered preferences.
-
Sep 2026New preprint: Remember Before You're Asked: MemDream for Self-Probing Memory Evolution — agents proactively probe and repair their memory before retrieval failures occur.
-
Aug 2026Our paper SEER: Long-Context Reasoning via Selective Visual-Text Compression is accepted to COLM 2026 🎉
-
Joined Tencent Hunyuan's Foundation Model Department as a Qingyun Intern.
-
2026New preprint: Beyond Sparse Supervision: Diffusion-Guided Learning for Few-Shot Graph Fraud Detection — diffusion-guided learning for few-shot graph fraud detection.
-
2026Our paper From Query to Counsel: Structured Reasoning with a Multi-Agent Framework and Dataset for Legal Consultation is accepted to ACL 2026 (Oral) 🎉
-
2026New preprint: Choosing How to Remember: Adaptive Memory Structures for LLM Agents — a framework (FluxMem) that lets agents pick the right memory structure per context. +9.18% / +6.14% on PERSONAMEM / LoCoMo.
-
2025From Newborn to Impact: Bias-Aware Citation Prediction accepted to The Web Conference 2026 🎉
-
2025Joined UTS as a PhD student under Prof. Yi Zhang. Research focus: agent memory and multi-agent reasoning.
📝 Selected Publications
PairPref: When Should Memory Guide the Answer? A Benchmark for Contextual Preference Use
arXiv preprint arXiv:2609.34526, 2026.
PairPref evaluates whether assistants adapt remembered preferences to changing situations. It pairs contexts in which the same preference should be applied or set aside, covering 1,227 pairs, 45 preferences, and eight situation categories. Reply selection and free generation reveal persistent difficulty with contextual preference use.
Remember Before You're Asked: MemDream for Self-Probing Memory Evolution
arXiv preprint arXiv:2609.34545, 2026.
MemDream proactively tests and evolves agent memory through offline dream cycles. Three agents — Dreamer, Analyst, and Consolidator — probe, diagnose, and repair the memory graph. A learned repair policy and reversible forgetting improve retrieval before real queries expose weaknesses, with +4.5 F1 points on LoCoMo and a +9.1-point overall score on MemoryAgentBench over the strongest reactive-evolution baselines.
From Query to Counsel: Structured Reasoning with a Multi-Agent Framework and Dataset for Legal Consultation
Annual Meeting of the Association for Computational Linguistics (ACL Main), 2026. CCF-A
We construct JurisCQAD, a large-scale, expert-verified dataset of 43,000 real-world Chinese legal consultation entries, and propose JurisMA, a modular multi-agent framework based on legal element graphs that supports dynamic routing, statute alignment, and style optimization. JurisMA significantly outperforms both general-purpose and legal-domain LLM baselines on LawBench.
From Newborn to Impact: Bias-Aware Citation Prediction
Proceedings of the ACM Web Conference (WWW), 2026, pp. 7600–7609. CCF-A
A bias-aware citation prediction framework addressing the challenge of forecasting impact for newly published papers under sparse early citation signals and long-tail distributions. By combining multi-agent feature extraction with robust graph representation learning, the model reduces MALE / RMSLE by ~13% and improves NDCG by 5.5%.
Choosing How to Remember: Adaptive Memory Structures for LLM Agents
arXiv preprint arXiv:2602.14038, 2026.
Existing agent memory systems use a one-size-fits-all structure and don't treat structure selection as a context-adaptive decision. We propose FluxMem, which equips agents with multiple complementary memory structures and learns to pick among them per interaction, with a three-level hierarchy and a Beta-Mixture-Model probabilistic gate replacing brittle similarity thresholds. +9.18% on PERSONAMEM, +6.14% on LoCoMo.
For the live citation count, see my Google Scholar.
📖 Education
-
Jan 2025 — presentPh.D. in Artificial Intelligence, University of Technology Sydney
Australian Artificial Intelligence Institute (AAII) · Advisor: Prof. Yi Zhang
👩🏫 Teaching
-
Feb 2025 — presentAcademic Tutor · Computer Science Studio 1 (41078)
Faculty of Engineering and Information Technology, University of Technology Sydney
🤝 Academic Service
Reviewer for: