Research Program

Agentic intelligence for complex, structured, and temporal data

My research studies how large language models can reason, act, and learn in data-rich environments. The central thread is LLM-driven agentic AI, with a long-running focus on time series analysis and broader applications in table mining, scientific literature mining, and recommender systems.

Core Direction
LLM Agents
Long-term Domain
Time Series
Research Style
Models + Systems
Evaluation Focus
Real Workflows
Representation Learning
Learning transferable signals from temporal, tabular, textual, and multimodal data.
Agentic Reasoning
Designing agents that use tools, memory, feedback, and multi-step reasoning.
Process-aware RL
Optimizing agent behavior with step-level and sequence-level reward signals.
Benchmarks & Systems
Building reusable platforms and evaluation tasks grounded in real research workflows.
LLMs and Agentic AI

This direction focuses on intelligent systems that reason, act, and improve through interaction. We study autonomous learning for large language models, slow-thinking reasoning, process-aware reinforcement learning, tool-augmented execution, memory, and agent runtime mechanisms. The goal is to move from prompt-only behavior toward agents that can plan, recover, and learn from environmental feedback in real workflows.

Agentic RL Slow Thinking Chain-of-Thought Tool Use Agent Memory Self-Evolution RAG LLM Evaluation
Time Series Analysis

This direction develops context-aware time series intelligence across representation learning, forecasting, classification, and anomaly detection. Earlier work builds neural architectures and self-supervised pre-training methods for temporal signals; recent work connects time series with multimodal language modeling, slow-thinking LLM reasoning, and agentic forecasting systems that select tools, engineer features, and refine predictions through structured interaction.

Agentic Forecasting LLM for Time Series Self-supervised Learning Multi-modal Reasoning Anomaly Detection Classification Foundation Models
Table Mining

LLM-driven tabular data understanding, table reasoning, multi-table inference, and scientific table analysis. We explore how LLMs can be augmented with programmatic tools and slow-thinking strategies to tackle complex structured data tasks.

Table Reasoning LLM + Tools Multi-table ChemTable
Scientific Literature Mining

Literature retrieval, multimodal scientific document understanding, knowledge discovery, and evidence-grounded reasoning over scholarly content. We build agentic systems and benchmarks for evaluating tool-augmented reasoning on academic literature.

RAG Literature Search Agentic Retrieval Benchmark
Recommender Systems

Sequential modeling, one-class collaborative filtering, and LLM-driven reasoning for dynamic user preference understanding. Research covers behavior-level data augmentation, cross-domain recommendation, and LLM-based user simulation.

Sequential Rec LLM + RecSys Cross-domain RL for Rec
Research collections: LLMs and Agentic AI · Time Series Analysis
欢迎脚踏实地而又积极主动的本科生、研究生同学加入认知智能全国重点实验室 USTC-AGentic Intelligence(AGI)Group