Research Projects
Selected research systems, benchmarks, datasets, and open-source implementations.
Agentic Systems
Strategic reasoning, collaboration, personalization, and secure code agents.
TreeDebater
Strategic Planning and Rationalizing on Trees Make LLMs Better Debaters
TreeDebater uses rehearsal and debate-flow trees to anticipate attacks, track active arguments, and allocate limited speaking time for more persuasive LLM debates.
SusVibes
Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks
SusVibes benchmarks 187 real-world feature requests and reveals a sharp gap between functional correctness and security in code produced by leading coding agents.
TypedThinker
TypedThinker: Typed Thinking Improves Large Language Model Reasoning
TypedThinker predicts when to use deductive, inductive, abductive, or analogical reasoning and supplies targeted demonstrations to diversify LLM problem solving.
PerSE
Learning Personalized Alignment for Evaluating Open-ended Text Generation
PerSE infers a reviewer's preferences from an in-context profile to deliver interpretable, fine-grained evaluations of how well open-ended generations align with that individual.
SALAM
Learning from Mistakes via Cooperative Study Assistant for Large Language Models
SALAM pairs an LLM with a cooperative study assistant that records prior mistakes and retrieves tailored guidance to help the model avoid recurring reasoning errors.
AI for Science
Machine learning for molecular explanation, antibodies, and peptide discovery.
RLHEX
Global Human-guided Counterfactual Explanations for Molecular Properties via Reinforcement Learning
RLHEX combines a variational graph generator with reinforcement learning to produce global molecular counterfactual explanations aligned with human-defined principles.
EATLM & ATUE
On Pre-training Language Model for Antibody
EATLM and the ATUE benchmark examine how general and antibody-specific protein language models transfer across antibody tasks and where biologically informed pre-training helps.
LSSAMP
Accelerating Antimicrobial Peptide Discovery with Latent Structure
LSSAMP jointly models peptide sequences and secondary structures in a quantized latent space to generate antimicrobial candidates, two of which showed strong activity in wet-lab tests.
Text Summarization
Faithful, multilingual, graph-based, and cross-domain summarization.
CALMS
Contrastive Aligned Joint Learning for Multilingual Summarization
CALMS uses contrastive sentence ranking and sentence-aligned substitution to improve multilingual summarization in both high- and low-resource languages.
CNewSum
CNewSum: A Large-scale Chinese News Summarization Dataset with Human-annotated Adequacy and Deducibility Level
CNewSum provides 304,307 Chinese news documents with human-written summaries and test-set adequacy and deducibility annotations for diagnosing summarization systems.
HeterSumGraph
Heterogeneous Graph Neural Networks for Extractive Document Summarization
HeterSumGraph connects sentence nodes through finer-grained semantic units in a heterogeneous graph to enrich cross-sentence modeling for extractive summarization.
MULTI-SUM
Exploring Domain Shift in Extractive Text Summarization
MULTI-SUM defines summarization domains by data source, measures cross-domain generalization gaps, and provides a testbed for comparing four adaptation strategies.