Sikun Guo

郭司坤

PhD Candidate in Computer Science, University of Virginia

RealAI Lab, University of Virginia · Advisor: Prof. Aidong Zhang · Charlottesville, VA

Research areas: LLM reasoning, knowledge recruitment and in-context learning, inference-time learning, graph neural networks, AI for Science, and autonomous AI research systems.

Research

Research agenda

I study how foundation models reason, use context, and improve ideas in scientific settings. My current work focuses on reasoning with large language models, theories of in-context learning, inference-time learning, and foundation models for scientific discovery. A connected thread studies graph-based scientific structure for biomedical hypothesis generation.

Model knowledge use

How can a model use available evidence, context, and internal knowledge more reliably in scientific reasoning tasks?

In-context learning

What theories and empirical evidence explain when examples, questions, or context help a model activate the right internal knowledge?

Inference-time learning

How can optimization, adversarial feedback, or reinforcement learning improve ideation and reasoning quality at test time?

AI for Science

How can models generate, critique, evaluate, and improve research ideas and scientific hypotheses in ways that are useful to researchers?

How I frame the problem space

The main question running through my work is how to make model reasoning more reliable. In many scientific settings, a model must combine context, structured evidence, and learned patterns when forming an answer or research idea.

That perspective pushes me toward both analysis and method design: designing curiosity-driven questions, and using inference-time adversarial or reinforcement learning to improve the trajectory of generated ideas.

I also care about the surrounding infrastructure. Benchmarks, truthfulness evaluation, graph-based scientific representations, and agentic ideation-rating systems make model-generated research easier to inspect and improve instead of treating it as polished prose alone.

Current directions

Representative threads in the agenda

  • Curiosity-driven questioning and engine-agnostic LLM research ideation.
  • Inference-time adversarial and reinforcement learning for idea optimization.
  • Truthfulness, hallucination, and reliability in model-generated scientific hypotheses.
  • Temporal, semantic, and hypergraph modeling of evolving biomedical concepts.