Short Bio

Research Overview

My research focuses on synergistic intelligence: enabling intelligent systems to learn from interaction, combine complementary expertise, and reuse experience across tasks. I study these questions in LLM agents, multi-agent systems, and federated learning.

(1) Learning Signals & Efficient Training: I develop methods that help intelligent systems learn from imperfect feedback and distributed data. My work spans credit assignment for tool-using agents (PORTool, NeurIPS'26), federated optimization with partial client participation (FedAMD, ICML'23), and calibrating updates when clients perform different amounts of local training (FedaGrac, IEEE TPDS'23). I also study training under resource constraints through adaptive submodel selection (FIARSE, NeurIPS'24) and federated LLM fine-tuning without the full model (FedBiOT, KDD'24), alongside alignment with heterogeneous human preferences (FedBiscuit, ICLR'25).

(2) Multi-Agent Collaboration & Information Exchange: I study how agents identify relevant collaborators and exchange information that others can use effectively. This includes iterative specialist routing for question answering (RIRS), latent-space communication across heterogeneous foundation models (Vision Wormhole), personalized knowledge transfer across different model architectures (KT-pFL, NeurIPS'21), and knowledge sharing across sensing modalities (FedHKD, SenSys'24). To make collaboration more efficient, I investigate sparse model communication (GlueFL, MLSys'23) and sign-based synchronization (Marsit, DAC'22). Related work uses probabilistic masking to mitigate input reconstruction risks and reduce communication in split federated learning (PM-SFL, KDD'26 Oral).

(3) Agentic Discovery & Skill Reuse: I study how agents transform interaction experience into reusable algorithms and skills. My work includes evolving optimization algorithms that combine heuristics with mathematical solvers under runtime constraints (HeurEvo) and transferring learned web interaction patterns across structurally compatible websites (SkillMigrator, NeurIPS'26). These methods preserve useful experience in executable procedures, allowing new tasks to build on earlier search and learning.

I also explore Trustworthy AI & Evaluation, including evaluating and defending against verbatim reproduction of copyrighted text (SHIELD, EMNLP'24), selective unlearning that reduces such reproduction while preserving unrelated capabilities (SUV, COLM'25), and scalable factuality evaluation using knowledge graphs (GraphEval, IEEE Data Engineering Bulletin'24).

I am always happy to discuss research ideas and collaborations.