LLM agents and multi-agent systems
How a team of LLM agents is wired together can matter as much as the model behind each agent.
I study how to learn that structure with reinforcement learning, and how to give enterprise agents
realistic environments to learn in before they meet production systems.
-
Adaptive topology learning for LLM multi-agent systems
First author, under review at ICLR 2027
A reinforcement learning framework that autoregressively learns sparse, task-adaptive agent communication topologies from execution feedback and performs targeted, failure-aware repair. Reaches 73% Pass@1 on LiveCodeBench against 58% for the strongest baseline, and cuts end-to-end token consumption on GSM8K by 39%.
-
Synthesizing executable enterprise environments for LLM agents
Co-author, NeurIPS 2026
Automatically synthesizes executable replicas of closed-source enterprise applications from their documentation, with functional APIs, databases and agent tools. Training agents in these environments improved an 8B LLM’s task accuracy by 52%. I co-led the work and built the end-to-end web platform for environment generation.
Efficient and scalable graph learning
No single graph neural network wins on every graph, and the strongest ones rarely fit in GPU memory
on real data. I build methods that choose models adaptively and train within a fixed memory budget at web scale.
-
Self-Adaptive Graph Mixture of Models
First author, AAAI 2026
Learns, for every node, which graph neural network architectures to use and how to combine them, pruning weak ones during training. Matched or outperformed strong baselines on 16 benchmarks, with gains of up to 22% across link prediction, graph-level and node classification tasks.
-
Scalable and memory-efficient graph learning
First author, under review at AAAI 2027
GraphDDR, a divide, distill and route framework that combines structure-aware graph partitioning, multi-teacher distillation and sparse expert routing to scale GNN training to graphs with hundreds of millions of nodes and billions of edges. Up to 4.5× faster training with 89% lower GPU memory consumption.
-
Matrix sketching for fast graph neural network training
Co-author, TMLR 2026
Compresses the graph before training with Proper Orthogonal Decomposition. Trains on a 111M-node graph in under 8 GB of GPU memory with about 339× faster epochs.
Graph AI with clients
I translate client problems into graph learning systems and see them through from requirements to deployment.
-
Real-time fraud detection for a payment corporation in Vietnam
Proof of concept
Led the client engagement in Hanoi, working directly with senior leadership to define fraud scenarios, data requirements and success criteria. Spearheaded the project and authored the technical proposal for a real-time fraud detection system built on streaming graph AI.
-
Tax-evasion detection on billion-scale taxpayer graphs in Italy
Client deployment
Developed and deployed an end-to-end graph AI pipeline for a heterogeneous graph with billions of nodes, decoupled from the production graph database. Designed a graph transformer that handles label scarcity and high-degree hubs, using edge features for a 3% gain in downstream performance, and built an interactive visualization tool for million-node graphs.
Patents. Three applications filed through Fujitsu as primary inventor, on self-adaptive graph mixture of models,
memory-efficient large-graph training, and multi-agent LLM collaboration and failure repair.
Service. Peer reviewer for AAAI, KDD and ICLR.