Projects
Research projects and systems
This page highlights the artifacts and system-building layer behind the papers: benchmarks, method lines, and workflow-oriented research tools.
I do not think of research output as papers alone. The more interesting question is often what infrastructure, benchmark, workflow, or evaluation layer had to exist for the research to become convincing in the first place.
Knowledge recruitment failures in LLMs
LLM reasoning · 2025–2026
A research thread on why language models can contain relevant knowledge yet fail to recruit it into the answer during scientific reasoning.
- Uses causal evidence to study the gap between latent knowledge and usable reasoning behavior.
- Connects directly to broader questions about in-context learning ability and parametric knowledge use.
Research ideation benchmarks and optimization
Benchmarking and method line · 2024–2026
A line of work around measuring and improving LLM-based research idea generation, from IdeaBench to curiosity-driven questioning and inference-time learning.
- Includes IdeaBench, Curiosity-Driven Questioning, InfAL, and InfRL.
- Focuses on making ideation quality measurable, inspectable, and improvable without relying only on larger base models.
Foundation models for scientific discovery
Research workflow systems · 2024–Present
An ongoing agenda on using foundation models in scientific workflows: idea generation, critique, optimization, knowledge grounding, and hypothesis evaluation.
- Connects internal knowledge use with external scientific evidence and workflow-level evaluation.
- Includes work on truthfulness, hallucination, and the reliability of generated scientific claims.
Biomedical hypothesis generation with evolving structure
Graph-based scientific modeling · 2024–2025
A collaborative direction on representing biomedical knowledge with temporal, semantic, and higher-order structure for hypothesis generation.
- Includes HyHG, which received the ICDM 2025 Best Paper Award, and ConceptDrift in Bioinformatics 2025.
- Models scientific concepts as evolving structures rather than static keyword collections.
Agentic scientific ideation rating
Industry research system · Summer 2025
At Autoscience Institute, built an agentic LLM system for automatically rating scientific ideation quality.
- Translated research-ideation evaluation into a practical system workflow.
- Bridges academic work on ideation benchmarks with applied scientific-evaluation infrastructure.
Graph robustness and long-range dependency modeling
Graph ML · 2022–2023
Earlier work on robust graph neural networks under adversarial structural attacks and across different homophily regimes.
- Includes the GalNN direction from the M.S. thesis on adaptive aggregator selection and long-range dependency modeling.
- Builds on a broader graph-learning background that includes graph neural networks for IoT.
Cryogenic low-pass filters for quantum experiments
Undergraduate research and patent · 2018–2021
Nankai University research on low-pass filters based on ferromagnetic substance powder for cryogenic quantum experiments.
- Contributed to patent CN111540984B as the second inventor.
- Led an undergraduate research team that won first prize in the National Undergraduates’ Innovation and Entrepreneurship Training Program.