Agentic RAG and Deep Research: Retrieval-augmented generation (RAG), agentic search, and deep research agents that plan, retrieve, and synthesize evidence into long-form reports. We work on RAG-reasoning systems, multi-hop and open-domain question answering, knowledge grounding over knowledge bases and text, report-level evaluation, and safety guardrails for autonomous research pipelines.
Deep Research with Open-Domain Evaluation and Multi-Stage Guardrails for Safety (ACL 2026) · Towards Agentic RAG with Deep Reasoning: A Survey of RAG-Reasoning Systems in LLMs (EMNLP 2025) · DIVKNOWQA: Assessing the Reasoning Ability of LLMs via Open-Domain QA over Knowledge Base and Text (NAACL 2024) · Heterogeneous Knowledge Grounding for Medical Question Answering with Retrieval Augmented LLM (WWW 2024) · Evaluating Robustness of Generative Search Engines on Adversarial Factoid Questions (ACL 2024) · Benchmarking Multimodal RAG with Dynamic VQA and a Self-Adaptive Planning Agent (ICLR 2025)
Agentic Reinforcement Learning and LLM Post-Training: Reinforcement learning for language agents and post-training methods that scale reasoning: curriculum learning and problem curation, policy-improvement bandits, in-context learning at many-shot scale, self-generated demonstrations, self-evolving agent skills, and co-evolutionary verification.
Actor-Curator: Co-adaptive Problem Curation via Policy-Improvement Bandits for Scalable RL Post-Training (COLM 2026) · Many-Shot Scaling of In-Context Learning with Self-Generated Demonstrations (ACL 2026) · CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification (COLM 2026) · kNN-ICL: Compositional Task-Oriented Parsing Generalization with Nearest Neighbor In-Context Learning (NAACL 2024) · WebAgent-R1: Training Web Agents via End-to-End Multi-Turn Reinforcement Learning (EMNLP 2025) · Sample Lottery: Unsupervised Discovery of Critical Instances for LLM Reasoning (ICLR 2026) · TestNUC: Enhancing Test-Time Computing Approaches and Scaling through Neighboring Unlabeled Data Consistency (ACL 2025)
Generative Models for Structured Data and Time Series: Generative modeling beyond text: tabular data synthesis and imputation, diffusion models and diffusion language models, in-context example selection for data generation, missing-data recovery, and time series forecasting with latent state representations.
TabNAT: A Continuous-Discrete Joint Generative Framework for Tabular Data (ICML 2025) · DiffPuter: Empowering Diffusion Models for Missing Data Imputation (ICLR 2025) · TabGen-ICL: Residual-Aware In-Context Example Selection for Tabular Data Generation (ACL 2025) · From Observations to States: Latent Time Series Forecasting (ICML 2026) · Glocal Information Bottleneck for Time Series Imputation (NeurIPS 2025) · Locally Confident, Globally Stuck: The Quality-Exploration Dilemma in Diffusion Language Models (COLM 2026)
Graph Mining and Outlier Detection: Graph neural networks, graph anomaly and outlier detection, fraud detection, graph self-training and node selection, temporal and heterogeneous graph learning, graph data augmentation, and scalable open-source tooling for the community.
PyGOD: A Python Library for Graph Outlier Detection (JMLR 2024) · BANGS: Game-Theoretic Node Selection for Graph Self-Training (ICLR 2025) · TGTOD: A Global Temporal Graph Transformer for Outlier Detection at Scale (PAKDD 2025) · Data Augmentation for Supervised Graph Outlier Detection with Latent Diffusion Models (LoG 2024) · Enhancing Fairness in Unsupervised Graph Anomaly Detection through Disentanglement (TMLR 2025) · LEGO-Learn: Label-Efficient Graph Open-Set Learning (TMLR 2025) · Grad: Guided Relation Diffusion Generation for Graph Fraud Detection (WWW 2025, pp. 5308-5319)
Human-Agent Collaboration and Interactive Systems: Fully autonomous agents remain unreliable for many real tasks. We study mixed-initiative systems where humans and LLM agents share control: interaction protocols, human feedback and preference signals, interactive research-code development, agent-assisted scientific workflows, and peer-review support.
LLM-Based Human-Agent Collaboration and Interaction Systems: A Survey (ACL 2026) · RECODE-H: A Benchmark for Research Code Development with Interactive Human Feedback (ICLR 2026) · LLMs Assist NLP Researchers: Critique Paper (Meta-)Reviewing (EMNLP 2024) · Toward User Preference Alignment in LLM Recommendation via Explicit Context Feedback (IEEE CogMI 2025) · GLEAN: Active Generalized Category Discovery with Diverse LLM Feedback (EACL 2026)
LLM Agents and Agent Memory: How agents accumulate, organize, and reuse experience so they act coherently over long horizons and improve themselves over time. We work on agent memory architectures, hierarchical and graph-based memory, long-horizon and self-evolving agents, memory-action integration, personalization at test time, and tool learning.
A Survey of Agent Memory in the Second Half: Towards Self-Evolving and Long-Horizon Agents (TMLR 2026, Survey Certification) · GAM: Hierarchical Graph-based Agentic Memory for LLM Agents (ACL 2026) · PersonaAgent: Bridging Memory and Action for Personalized LLM Agents (ACL 2026) · CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification (COLM 2026)
Recommender Systems and Personalization: A long-standing strength of the lab, now driven by large language models and graph learning: LLM agents as recommenders, tool learning for recommendation, prompt refinement and reranking, graph-language token alignment, self-supervised and contrastive collaborative filtering, sequential and session-based recommendation, and cold-start settings.
AgentDR: Dynamic Recommendation with Implicit Item-Item Relations via LLM-based Agents (WWW 2026) · Personalized Recommendation Tool Learning via Autonomous Language Agents (RecSys 2026) · AGP: Auto-Guided Prompt Refinement for Personalized Reranking in Recommender Systems (WSDM 2026) · SGCL: Unifying Self-Supervised and Supervised Learning for Graph Recommendation (RecSys 2025) · LLMInit: A Free Lunch from Large Language Models for Selective Initialization of Recommendation (EMNLP 2025) · Training Large Recommendation Models via Graph-Language Tokens Alignment (WWW 2025) · Pre-Training with Transferable Attention for Cross-Market Sequential Recommendation (KDD 2024) · InterFormer: Heterogeneous Interaction Learning for Click-Through Rate Prediction (CIKM 2025)
Trustworthy LLMs: Hallucination, Safety and Evaluation: What makes these systems dependable. We study mechanistic analysis of hallucination over structured and linearized knowledge, watermarking and provenance, adversarial robustness and attacks, machine unlearning, multimodal safety, and benchmark and evaluation design for code and agent behavior.
Why LLMs Hallucinate on Structured Knowledge: A Mechanistic Analysis of Reasoning over Linearized Representations (ACL 2026) · MUSE: Model-Agnostic Tabular Watermarking via Multi-Sample Selection (ICLR 2026) · MultiFileTest: A Multi-File-Level LLM Unit Test Generation Benchmark (ACL 2026) · DA³: A Distribution-Aware Adversarial Attack against Language Models (EMNLP 2024) · Refine Knowledge of LLMs via Adaptive Contrastive Learning (ICLR 2025) · Jailbreaking LLMs Through Alignment Vulnerabilities in Out-of-Distribution Settings (CIKM 2025) · Reefknot: A Comprehensive Benchmark for Relation Hallucination Evaluation (ACL 2025) · PersonaGym: Evaluating Persona Agents and LLMs (EMNLP 2025)
Uncertainty Quantification and Conformal Prediction: Toward models that know when they should not answer. We develop conformal prediction for graphs and stream networks, distribution-free coverage guarantees, calibration and confidence estimation, uncertainty in graph neural networks, data selection under distribution shift, and deployment audits for high-stakes triage.
Topology-Aware Conformal Prediction for Stream Networks (NeurIPS 2025) · Uncertainty in Graph Neural Networks: A Survey (TMLR 2024) · A Deployment Audit of Release-Side Risk in Conformal Triage under Prevalence Shift (COPA 2026) · Confidence-Aware Fine-Tuning of Sequential Recommendation Systems via Conformal Prediction (IEEE BigData 2025) · SeSE: Black-Box Uncertainty Quantification for Large Language Models (UAI 2026)
Funded Projects: Exploiting the Massive User Generated Utterances for Intent Mining under Scarce Annotations (NSF 1909323) · Learning Dynamic and Robust Defenses Against Co-Adaptive Spammers (NSF 1930941)
Earlier directions of the lab, retained for reference.
Deep Learning: With the ever-increasing information we gathered from various information sources either online or offline, deep learning significantly contributes to a wide range of research tasks from which representative features can be learned and utilized with minimized hand-craft feature engineering. Our research on deep learning focuses on knowledge discovery/extraction and representation learning on user-generated data such as online text queries, user reviews as well as information/social networks.
Graph Mining: Graphs are increasingly important in modeling real-world data with complex structures. Our research on graph mining includes the following topics:
-Large graph database management: graph search and indexing on a database of large graphs, and on a single large network.
-Scalable machine mining on large graph(s): community detection, link inference, and collective classification.
-Subgraph pattern mining: finding and extracting useful information from graph structured data sets (e.g., molecular structure graphs) to discover significant features.
Healthcare: The human brain is one of the most complicated biological structures in the known universe. It is very challenging to understand how it works, especially when disorders and diseases occur. Multiple data representations are usually involved, including multi-view biomarkers, neuroimaging tensor data, brain network data, and sequential user behavior data. We collaborate with the CoNECt Lab on various neuroimaging and healthcare projects.
-Computer-aided diagnosis: We focus on fusing heterogeneous data sources to assist diagnosis.
-Precision medicine: We aim to leverage deep learning techniques to provide personized healthcare services.
-Mobile health: The increasing use of electronic forms of communication presents new opportunities in the study of healthcare, including the ability to investigate the manifestations of psychiatric diseases unobtrusively and in the setting of patients’ daily lives. We aim to study the connections between mood disorders and mobile phone usage.
A Review of Heterogeneous Data Mining for Brain Disorder Identification by Bokai Cao.
Heterogeneous Information Networks: Many real-world networks like social networks and information systems usually involve a large number of components, multiple types entities interconnected with different types of relations. We call these networks as heterogeneous information networks, which are critical for modern information infrastructure.
Learning from Multiple Data Sources: Multiple related data sources containing different types of features may be available for a given task. For instance, users’ profiles can be used to build recommendation systems; in addition, a model can also use users’ historical behaviors and social networks to infer users’ interests on related products. It is desirable to collectively use any available multiple heterogeneous data sources in order to build effective learning models, including transfer learning, crowd sourcing, and heterogeneous learning.
Mining Uncertain and Incomplete Data: Most real data we are facing these days are neither certain nor complete, which becomes a great challenge for applying conventional data mining methods on these data. We aim at designing effective models to perform knowledge discovery from data with uncertainty and incompleteness.
Multi-label Learning: Many real-world classification tasks involve multiple concepts instead of one single concept, and each data object can be assigned with multiple concepts (class labels) simultaneously. Multi-label learning aims at building accurate classification models that can predict multiple concepts collectively for each object.
Natural Language Processing: Exploiting the Massive User Generated Utterances for Intent Mining under Scarce Annotations
Outlier Detection: A Python Library for Graph Outlier Detection (PyGOD)
Privacy Preserving Data Publishing: Privacy-preserving data publishing provides methods and tools for publishing useful information while preserving data privacy.
Recommender System: With the ability to provide personalized suggestions, recommender systems have become an important tool in many web services for attracting and retaining users. We focus our research on building the next generation of recommender systems which can better understand people’s needs. In particular, we aim to utilize deep learning to alleviate the cold-start problem which is a common issue in recommender systems.
A Survey and Critique of Deep Learning on Recommender Systems by Lei Zheng.
Social Network Analysis: Powered by data cloud and mapreduce infrastructure, social network platforms are gathering data on many aspects of our daily lives. Motivated by this trend, our research addresses interesting phenomena on social networks including the following topics:
-Network structure and macro social pattern mining: magnet community detection, and social influence evaluation.
-Influence propagation and social activity mining: social sharing temporal pattern mining, spam detection, and social advertising.
-Role discovery: finding the most influential nodes.
Link Prediction across Heterogeneous Social Networks: A Survey by Jiawei Zhang.
Spam Detection: Learning Dynamic and Robust Defenses Against Co-Adaptive Spammers
Stream Mining: Design efficient real-time algorithms for continuous data streams, especially for graph streams.