EvoDS: Self-Evolving Autonomous Data Science Agent with Skill Learning and Context Management
A self-evolving data-science agent that acquires reusable executable skills and learns when to compress long-horizon context.
Selected publications
Agentic Data Scientists
MM-Agent: LLM as Agents for Real-world Mathematical Modeling Problem
An end-to-end agent that translates open-ended real-world problems into mathematical models, executable solutions, and structured reports.
Towards Multimodal Data-Driven Scientific Discovery Powered by LLM Agents
MoSciBench evaluates whether agents can align heterogeneous scientific data, execute analysis pipelines, and verify hypotheses across modalities and domains.
Foundation Models for Scientific Discovery: From Paradigm Enhancement to Paradigm Transition
A three-stage view of foundation models evolving from scientific tools, to human-AI collaborators, and ultimately to autonomous discovery agents.
Grounded Physical Agents
DeepTravel: An End-to-End Agentic Reinforcement Learning Framework for Autonomous Travel Planning Agents
An end-to-end agentic reinforcement-learning framework that trains travel-planning agents through a realistic tool sandbox, hierarchical feedback, and experience replay.
USTBench: Benchmarking and Dissecting Spatiotemporal Reasoning of LLMs as Urban Agents
A benchmark and interactive environment that tests how urban agents understand, forecast, plan, and learn from feedback in spatiotemporal worlds.
CoLLMLight: Cooperative Large Language Model Agents for Network-Wide Traffic Signal Control
A cooperative LLM-agent system that reasons over network-wide traffic interactions and uses simulation feedback to improve coordinated control.
An LLM-Powered Cooperative Framework for Large-Scale Multi-Vehicle Navigation
A hierarchical system in which a global agent allocates network-wide traffic flow and local agents make adaptive, congestion-aware routing decisions.
DiMA: An LLM-Powered Ride-Hailing Assistant at DiDi
A deployed ride-hailing assistant that combines tool-augmented spatiotemporal order planning, cost-aware dialogue generation, and continual fine-tuning.
LLMLight: Large Language Models as Traffic Signal Control Agents
An early demonstration that language models can turn structured traffic observations into interpretable reasoning and closed-loop signal-control actions.
UrbanKGent: A Unified Large Language Model Agent Framework for Urban Knowledge Graph Construction
A tool-augmented agent family for urban relation extraction and knowledge-graph completion, built through knowledgeable instructions and iterative trajectory refinement.
Hydra: A Personalized and Context-Aware Multi-Modal Transportation Recommendation System
A large-scale decision system that combines user preferences, route plans, and heterogeneous urban context for personalized multi-modal transportation recommendation.
Models of Dynamic Worlds
UniExtreme: A Universal Foundation Model for Extreme Weather Forecasting
A universal foundation model that learns event-aware representations for transferable forecasting across diverse extreme-weather conditions.
Scalable Pre-Training of Compact Urban Spatio-Temporal Predictive Models on Large-Scale Multi-Domain Data
A compact pre-training framework that learns transferable urban dynamics across heterogeneous domains through mixture normalization, multi-scale mixing, and adaptive tuning.
BigST: Linear Complexity Spatio-Temporal Graph Neural Network for Traffic Forecasting on Large-Scale Road Networks
A linear-complexity spatiotemporal graph network that captures long-range dynamics while scaling traffic forecasting to very large road networks.
Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach
A transformable patching architecture that regularizes irregular observations while learning time-varying dependencies across multivariate scientific time series.
iETA: A Robust and Scalable Incremental Learning Framework for Time-of-Arrival Estimation
A production-scale ETA framework that continuously adapts to evolving traffic through incremental prediction, knowledge consolidation, and adversarial training.