Research Work
Selected work on agent memory, multi-agent orchestration, and LLM evaluation, followed by earlier research projects.
T-Mem: Memory That Anticipates, Not Archives
2026EMNLP 2026 Main Conference
- Developed long-term conversational memory that uses write-time triggers to support both descriptive and associative recall.
- Connected this research to practical memory retrieval and response-quality evaluation during my Tencent internship.
Long-Term MemoryAssociative RetrievalPersonalized Agents
AMA: Adaptive Memory via Multi-Agent Collaboration
2026Findings of ACL 2026
- Developed a collaborative agent framework for multi-granularity memory construction, adaptive retrieval, consistency checks, and targeted updates.
- Evaluated long-context interaction while reducing token use compared with full-context approaches.
Agent MemoryMulti-Agent CollaborationLifelong Learning
Unified-MAS: Domain-Specific Nodes for Automatic Multi-Agent Systems
2026EMNLP 2026 Main Conference
- Generated specialized nodes using external domain knowledge and optimized bottleneck nodes before system orchestration.
- Improved the performance-cost trade-off of automatic multi-agent systems across knowledge-intensive domains.
Multi-Agent OrchestrationNode GenerationLLM Agents
BizCompass: LLM Reasoning in Business Knowledge and Applications
2026Findings of ACL 2026
- Helped build a benchmark connecting business knowledge in finance, economics, statistics, and operations to application-level reasoning tasks.
- Analyzed how foundational capabilities support performance in practical business roles.
LLM EvaluationBenchmarkReasoning

UNGIA: Unnoticeable Injection Attack on Graph Neural Networks
2026IJCNN 2026
- Investigated stealthy graph injection attacks on graph neural networks.
- Developed UNGIA, a method that selects representative nodes and uses bi-level optimization to balance homogeneity and attack effectiveness.
- Conducted experiments on Cora, Citeseer, and Cora-ML benchmark datasets.
Graph Neural NetworksSecurityAdversarial Learning

CNO-former and Chaos-Based Neural Models
2023 - 2024Selected undergraduate research projects
- Formulated CNO-former, a Transformer-style model that modifies neuron states within self-attention to reduce computational cost and mitigate information loss.
- Explored chaotic neural oscillation mechanisms for aspect-based sentiment analysis.
- Awarded an Excellent Award in the 2nd Computer Science Related Poster Exhibition.
TransformerSelf-AttentionSentiment Analysis