Research Areas
Exploring the frontiers of large language models, brain-inspired AI, and self-evolving agents
My research focuses on advancing large language model capabilities through seven interconnected areas: brain-inspired memory architectures, multi-agent systems, retrieval-augmented generation, reasoning mechanisms, domain-specific LLMs, code generation, and — as the current doctoral direction at PolyU — recursive self-improvement. The goal is to develop AI systems that collaborate effectively, retain long-term memory, reason through complex problems, and ultimately turn their own experience into lasting capability.
Brain-Inspired Memory Architecture
FeaturedThis research draws on neuroscience principles to develop memory systems for multi-agent AI frameworks. The BMAM (Brain-inspired Multi-Agent Memory) framework models key brain regions — hippocampus for spatial/episodic memory, prefrontal cortex for reasoning and personality, and temporal lobe for semantic processing — to enable agents with human-like memory consolidation, forgetting, retrieval, and reflection capabilities.
BMAM Architecture Overview
Key Contributions
- BMAM Framework — Multi-agent memory system with brain region agents (Findings of ACL 2026, arXiv:2601.20465)
- Memory consolidation and distortion modeling
- Personality-driven reasoning validation
Publication
Y. Li, J. Liu, Y. Wang, Y. Wu, and M. Xu. “BMAM: Brain-inspired Multi-Agent Memory Framework.” Findings of ACL 2026. arXiv:2601.20465
Multi-Agent Systems
This research area focuses on developing frameworks for multiple AI agents to collaborate through specialization and coordination. I investigate both the design of agent architectures and evaluation methodologies to measure effectiveness in complex task environments.
Current Projects
- XAgent evaluation framework development (Tsinghua NLP Lab)
- Inter-agent communication protocols
- Specialized agent roles and coordination mechanisms
Key Collaborations
Tsinghua University NLP Lab, GDIIST
Retrieval-Augmented Generation (RAG)
My work on RAG focuses on enhancing LLMs by retrieving and effectively utilizing external knowledge. The ReproAgent project aims to improve retrieval precision, context integration, and query formulation techniques for more reproducible and robust AI agent research.
Current Projects
- ReproAgent: Reproducible agents through enhanced RAG (Tsinghua NLP Lab)
- Context integration techniques for complex queries
- Domain-specific retrieval optimization for medical and scientific domains
Key Collaborations
Tsinghua University NLP Lab, HKMU Research Team
Reasoning in LLMs
EMNLP 2025This research investigates methods to enhance reasoning capabilities in LLMs, particularly for handling ambiguous questions. The CondAmbigQA benchmark (published at EMNLP 2025 Main Conference) demonstrates that condition reasoning improves LLM answer accuracy by up to 18.9%, reframing apparent hallucinations as query ambiguity.
Key Contributions
- CondAmbigQA: 2,000-query benchmark with condition-aware evaluation metrics
- Demonstrated 18.9% accuracy improvement via condition reasoning
- Contextual disambiguation techniques
Publication
Z. Li*, Y. Li*, H. Xie, and S. J. Qin. “CondAmbigQA: A Benchmark and Dataset for Conditional Ambiguous Question Answering.” EMNLP 2025 (Main Conference), pp. 2269-2288. *Equal contribution
Domain-Specific LLMs
CurrentBuilding specialized large language models for pain analysis and assessment (Pain LLM) at GDIIST. This research involves multi-node distributed training on 20+ NVIDIA H100 GPUs, developing pain mouse behavioral benchmarks, and contributing to the China Brain Project.
Training Infrastructure
Current Projects
- Pain LLM — Domain-specific model for pain analysis and assessment
- Pain Mouse Benchmark — Behavioral evaluation for pain-related models
- Contributions to the China Brain Project
- FDCT (Macao Science and Technology Development Fund) grant proposals
Institution
Guangdong Institute of Intelligence Science and Technology (GDIIST)
Code Generation
KDD 2023Contributed to CodeGeeX, a pre-trained model for code generation with multilingual benchmarking on HumanEval-X. Developed GLM-based algorithms for poetry generation and content optimization, and improved code-generation models through data analysis and prompt engineering at Zhipu AI (THU KEG).
Key Contributions
- CodeGeeX — Multilingual code generation model (KDD 2023)
- GLM-based poetry generation and content optimization
- Prompt engineering for code generation improvement
Publication
Q. Zheng, X. Xia, et al. “CodeGeeX: A Pre-Trained Model for Code Generation with Multilingual Benchmarking on HumanEval-X.” KDD 2023
Recursive Self-Improvement & Self-Evolution
PhD DirectionToday's agents are frozen between runs: whatever they learn while solving a task is discarded when the context ends. My doctoral direction at PolyU asks what changes when that loop is closed — when an agent's own experience feeds back into the strategies, prompts, and tools it uses next time, and those revisions accumulate. This is where the memory work and the agent work meet: memory is the substrate that makes self-improvement possible, and the structure of that memory largely decides whether improvement compounds or drifts.
The Self-Evolution Loop (Sketch)
A directional sketch, not an architecture. Concrete designs and results are kept for the papers.
Open Questions I'm Pursuing
- Convergence vs. drift — what property of the loop makes repeated self-revision improve rather than degrade a system?
- Measuring a moving target — how do you evaluate improvement when the evaluator, the memory, and the policy all change together?
- Credit over long horizons — which past experience deserves to be kept, and which should be allowed to fade?
- Bounded self-modification — what verification envelope keeps self-directed change reversible, interpretable, and safe?
- Open-ended evaluation — benchmarks that do not saturate as the agent gets better at them.
Relation to Other Work
Continues directly from the brain-inspired memory line (BMAM) and the multi-agent line above, and sits on Choice Point 04 of the Star Trail route — the fork between unconstrained and supervised recursive growth.
Status
Early-stage doctoral research. Described here at the level of direction; specific methods, systems, and preliminary results are intentionally not detailed until publication.