Preprint ·
TimeRSI: Towards Self-Improved Agentic Deployment for Time-Series Foundation Models
A coding agent improves deployment pipelines around frozen time-series foundation models using historical forecast errors and adaptive pipeline selection.
Research Projects
Research across multimodal foundation models, LLM agents, post-training, and responsible AI. See the full publication list and Google Scholar.
Preprint ·
Can Language Models Learn to Reject Their Own Bad Reasoning Steps?
Self-Step Rejection trains a lightweight acceptance gate on a frozen language model to reject harmful reasoning steps and resample within a bounded budget.
Preprint ·
From Knowledge Access to Source Learning: Developing Source-Specific Competence
SourceLearn builds reusable, source-specific competence through self-directed study and task-guided refinement of a persistent source model.
Preprint ·
Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems
SciUtopia simulates academic research ecosystems with persistent LLM agents, connecting research choices, collaboration, peer review, funding, and citations over time.
NeurIPS'26 Workshop ·
Beyond Structural Unification: A Perspective and Roadmap on Unified Multimodal Models
A position paper outlining five layers of functional unification for multimodal understanding and generation: inputs, representations, reasoning, optimization, and evaluation.
EMNLP'26 ·
CultureVLM: Characterizing and Improving Cultural Understanding of Vision-Language Models for over 100 Countries
CultureVLM characterizes and improves cultural understanding of vision-language models across more than 100 countries using culturally-grounded benchmarks and training procedures.
EMNLP'26 Main ·
MASCOT: Multi-Agent Socio-Collaborative Companion Systems
MASCOT is a multi-agent framework for socio-collaborative companions that uses bi-level optimization—persona-aware behavioral alignment and collaborative dialogue optimization—to counter persona collapse and social sycophancy, improving role consistency and reducing redundant dialogue.
ACL'26 Industry ·
MM-BizRAG: Rethinking Multimodal Retrieval-Augmented Generation for General Purpose Enterprise Q&A
Structure-aware multimodal RAG for enterprise Q&A, combining layout-aware report parsing, page-level slide representations, and inference-time context assembly. Includes FastRAGEval for efficient answer evaluation.
ACL'26 ·
Reasoning Is Not All You Need: Examining LLMs for Multi-Turn Mental Health Conversations
A systematic study showing that reasoning capabilities alone are insufficient for LLMs in multi-turn mental health conversations, isolating failure modes that demand additional safety and empathy-aware design.
ACL'26 · Oral ·
SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
SARA is a unified RAG framework that balances local factual precision with global coverage by combining natural-language spans with compact semantic compression vectors, achieving consistent gains under strict context budgets.
ICWSM'26 ·
MedHalu: Hallucinations in Responses to Healthcare Queries by Large Language Models
MedHalu is a fine-grained benchmark for studying hallucinations in LLM responses to consumer healthcare queries, analyzing hallucination patterns across models, query types, and medical specialties.
COLM'26 Efficient Reasoning Workshop ·
TextReg: Mitigating Prompt Distributional Overfitting via Regularized Text-Space Optimization
Regularizes prompt optimization to limit prompt bloat and narrow, sample-specific rules. Combines gradient purification, semantic edit regularization, and regularization-guided updates to improve out-of-distribution generalization.
COLM'26 LLA Workshop ·
UniSD: Towards a Unified Self-Distillation Framework for Large Language Models
A unified study of LLM self-distillation that combines teacher agreement, EMA stabilization, contrastive learning, feature matching, and divergence clipping to improve adaptation without stronger external teachers.
ICLR'26 ·
Beyond Magic Words: Sharpness-Aware Prompt Evolving for Robust Large Language Models with TARE
Sharpness-aware prompt evolution optimizes prompts for both performance and robustness by penalizing sharp regions in the prompt loss landscape, yielding prompts that transfer better across tasks and LLM families.
ICLR'26 ·
Sysformer: Safeguarding Frozen Large Language Models with Adaptive System Prompts
Sysformer learns adaptive, query-conditioned system prompts to safeguard frozen large language models, providing fine-grained safety control without modifying model weights.
NeurIPS'26 ·
AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent
Distills multi-agent debate into one LLM through reasoning-enhanced fine-tuning, trajectory-based augmentation, and process-aware distillation, moving computation from inference to training.
Preprint ·
Consistency Should Be the Priority for Unified Multimodal Models
A position paper on consistency across multimodal grounding, representations, reasoning, post-training, and evaluation.
ACL'26 ·
SlideAgent: Hierarchical Agentic Framework for Multi-Page Visual Document Understanding
We introduce SlideAgent, a versatile agentic framework for understanding multi-modal, multi-page, and multi-layout documents.
EMNLP'25 Findings ·
Protein Large Language Models: A Comprehensive Survey
A comprehensive survey of Protein Large Language Models, organizing the field along architectures, training objectives, datasets, and downstream tasks across biology, chemistry, and medicine.
KDD'25 TGL Workshop ·
Empowering Interdisciplinary Insights with Dynamic Graph Embedding Trajectories
DyGET converts temporal graph embeddings into interpretable trajectories that empower interdisciplinary insights through cross-disciplinary discovery and longitudinal community analysis.
AI and Ethics ·
Deconstructing the ethics of large language models from long-standing issues to new-emerging dilemmas: a survey
A survey deconstructing the ethics of large language models — from long-standing issues such as bias and misinformation to newly emerging dilemmas around agentic behavior and cross-cultural alignment.
Preprint ·
Efficient Knowledge Probing of Large Language Models by Adapting Pre-trained Embeddings
An efficient approach to probing LLM knowledge that adapts pre-trained embeddings to query model knowledge with substantially reduced compute.
ACL'25 ·
A Survey on Efficient Large Language Model Training: From Data-centric Perspectives
A survey of efficient LLM training organized around data-centric techniques — selection, mixing, ordering, and synthesis — and their trade-offs with compute and downstream performance.
ICML'25 FM4LS Workshop ·
ProteinGPT: Multimodal LLM for Protein Property Prediction and Structure Understanding
ProteinGPT is a multimodal LLM that integrates protein sequence and structural representations in a unified generative interface for property prediction and structure understanding.
Preprint ·
Topological Structure Learning Should Be A Research Priority for LLM-Based Multi-Agent Systems
We propose a framework for developing topology-aware Multi-Agent Systems (MAS), emphasizing agent selection, structure profiling, and topology synthesis, to enhance coordination and efficiency in complex task.
Preprint ·
A Comprehensive Survey of Evaluating Multimodal Foundation Models: Hierarchical Perspective and Extensive Applications
A survey of multimodal foundation model evaluation from a hierarchical perspective, covering a broad range of applications.
WebConf'25 Companion ·
ScreenLLM: Stateful Screen Schema for Efficient Action Understanding and Prediction
ScreenLLM introduces a stateful screen schema and key-frame extractor that compresses dynamic UI sessions into time-aware summaries, enabling efficient GUI understanding and action prediction with multimodal LLMs.
Preprint ·
The Influence of Text Variation on User Engagement in Cross-Platform Content Sharing
Studies how rewriting YouTube video titles when sharing them on Reddit affects engagement, using controlled comparisons to isolate the effects of language and community context.
Preprint ·
Large Language Model Agent: A Survey on Methodology, Applications and Challenges
A methodology-centered survey of LLM agents, connecting agent architectures, collaboration, and evolution with evaluation, tools, applications, and open challenges.
AAAI'25 Workshop · Best Paper ·
SciEvo: A 2 Million, 30-Year Cross-disciplinary Dataset for Temporal Scientometric Analysis
SciEvo is a comprehensive dataset containing 2 million+ papers spanning 30 years (1995-2024) for temporal scientometric analysis.
AAAI'25 DAI Workshop ·
UniGuard: Towards Universal Safety Guardrails for Jailbreak Attacks on Multimodal Large Language Models
UniGuard is a universal safety guardrail for multimodal LLMs, defending against cross-modal jailbreak attacks across image and text channels with low utility cost.
NeurIPS'24 MLSB Workshop ·
RNA-GPT: Multimodal Generative System for RNA Sequence Understanding
RNA-GPT is a multimodal generative system that combines RNA sequence reasoning with structural cues for property prediction, retrieval, and natural-language interaction over RNA data.
EMNLP'24 · Oral ·
AgentReview: Exploring Peer Review Dynamics with LLM Agents
Peer review is fundamental to the integrity and advancement of scientific publication. Traditional methods of peer review analyses often rely on exploration and statistics of existing peer review data...
EMNLP'24 ·
Large Language Models Can Be Contextual Privacy Protection Learners
PrivacyMind teaches LLMs to be contextual privacy protection learners that recognize sensitive content in context and adapt outputs accordingly, preserving utility while reducing leakage.
CIKM'24 ·
Towards Fair Graph Anomaly Detection: Problem, Benchmark Datasets, and Evaluation
We address fairness issues in graph anomaly detection, providing benchmark datasets and comprehensive evaluation frameworks for fair anomaly detection on graphs....
ICML'24 · Oral ·
CompeteAI: Understanding the Competition Dynamics of Large Language Model-based Agents
This work studies the competition dynamics among LLM-based agents, revealing emergent behaviors and strategic patterns in multi-agent systems....
WebConf'24 · Oral ·
Better to Ask in English: Cross-Lingual Evaluation of Large Language Models for Healthcare Queries
We present a framework and benchmark to evaluate LLMs' multilingual capabilities in healthcare queries, revealing significant performance gaps across languages and providing insights for improving hea...
ACL'24 ·
MM-Soc: Benchmarking Multimodal Large Language Models in Social Media Platforms
Social media platforms are hubs for multimodal information exchange, encompassing text, images, and videos, making it challenging for machines to comprehensively understand the information. Multimodal...
ACL'24 · Oral ·
Prototypical Reward Network for Data-Efficient RLHF
We propose a prototypical reward network that enables data-efficient reinforcement learning from human feedback (RLHF) for large language models....
ICML'23 ·
Semi-Offline Reinforcement Learning for Optimized Text Generation
We propose a semi-offline reinforcement learning approach for optimizing text generation in language models, balancing exploration and exploitation effectively....
WebConf'23 · Oral ·
Code Recommendation for Open Source Software Developers
CODER is a graph-based code recommendation framework for open source developers, modeling heterogeneous OSS contribution networks to deliver multi-modal recommendations across realistic workloads.
KDD'23 · Oral ·
Predicting Information Pathways Across Online Communities
We develop methods to predict how information spreads across different online communities, revealing patterns in cross-platform information diffusion....
AAAI'23 · Oral ·
Prototypical Fine-tuning: Towards Robust Performance Under Varying Data Sizes
We propose prototypical fine-tuning, a novel framework for fine-tuning pretrained language models that maintains robust performance across varying data sizes.
KDD'22 · Oral ·
Reinforcement Subgraph Reasoning for Fake News Detection
A novel subgraph reasoning paradigm for fake news detection that provides explainability while improving generalization through reinforcement learning and hierarchical graph attention networks.
AAAI'22 · Oral ·
Towards Fine-Grained Reasoning for Fake News Detection
We propose a fine-grained reasoning framework for fake news detection by following the human information-processing model and designing a prior-aware bi-channel kernel graph network....