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.
LLM reasoning · in-context learning · inference-time learning · AI for Science
My research asks how large language models can reason more reliably in scientific settings. I study in-context learning and inference-time optimization so foundation models can better generate, critique, and refine research ideas and hypotheses. The broader goal is to make AI systems more capable scientific collaborators: grounded, inspectable, and useful in real research workflows.
Current position, research agenda, and recent work
Reasoning with large language models
In-context learning and model knowledge use
Inference-time learning for research ideation
Foundation models for scientific discovery
Graph neural networks and structured scientific knowledge
Research agenda
I work on a single underlying question: how can foundation models reason more reliably with the information available to them? That question appears in different forms—curiosity-driven ideation, inference-time optimization, truthfulness evaluation, and graph-based scientific modeling—but the agenda is consistent.
How can language models use available information more reliably when generating scientific ideas, hypotheses, and answers?
What makes examples, questions, and context activate better scientific reasoning from the same underlying model?
How far can adversarial feedback or reinforcement learning improve ideation, critique, and reasoning quality at test time?
What makes model-generated ideas, hypotheses, and scientific claims trustworthy, checkable, and worth a researcher’s time?
Short bio
I am Sikun Guo (郭司坤), a PhD Candidate in Computer Science at the University of Virginia in the RealAI Lab, advised by Prof. Aidong Zhang. My research studies reasoning with large language models, theories of in-context learning, inference-time learning, and foundation models for scientific discovery. Recent work includes curiosity-driven questioning for research ideation, inference-time adversarial and reinforcement learning for idea optimization, truthfulness evaluation for scientific hypotheses, and graph-based biomedical discovery. Outside research, I follow financial markets and trading, play table tennis, sing, and enjoy road trips.
Selected publications
A representative slice of the research program, spanning curiosity-driven ideation, inference-time learning, scientific reasoning evaluation, and graph-based discovery.
Sikun Guo, Di Wang, Xiaohan Fan, Albert Huang and Aidong Zhang
KDD 2026
Studies whether curiosity-oriented questioning can make research ideation with language models more exploratory, engine-agnostic, and practically useful. The method emphasizes question generation, filtering, and clarity scoring before ideas are handed to downstream ideation engines.
Sikun Guo, Amir Hassan Shariatmadari, Jiuqi Wang, Albert Huang, Stefan Bekiranov, Shangtong Zhang and Aidong Zhang
KDD 2026
Explores inference-time reinforcement learning as a way to optimize the quality of model-generated research ideas without relying only on larger base models. The project frames idea improvement as a trajectory with actions, rewards, and iterative policy updates at test time.
Amir Hassan Shariatmadari, Sikun Guo, Nathan C. Sheffield, Aidong Zhang and Kishlay Jha
IEEE ICDM 2025
ICDM 2025 Best Paper Award
HyHG models biomedical hypothesis generation with a temporal hypergraph view of evolving scientific concepts, allowing the system to capture higher-order relations that ordinary pairwise graphs miss. The paper received the ICDM 2025 Best Paper Award, making it a particularly visible result in the structured-scientific-discovery line of work.
Sikun Guo, Amir Hassan Shariatmadari, Peng Wang, Albert Huang and Aidong Zhang
Findings of EMNLP 2025
InfAL studies how adversarial learning at inference time can improve the quality of model-generated research ideas without changing the base model weights. The paper pushes research ideation beyond prompt engineering toward explicitly optimized test-time behavior.
Sikun Guo, Amir Hassan Shariatmadari, Guangzhi Xiong, Albert Huang, Myles Kim, Corey Williams, Stefan Bekiranov and Aidong Zhang
KDD 2025
IdeaBench introduces a benchmark and evaluation framework for LLM-based research idea generation. The core contribution is not only a dataset, but also a way to measure whether generated ideas are insightful, plausible, and meaningfully distinct rather than merely well phrased.
Autoscience Institute · Summer 2025
Automated scientific ideation rating with LLMs by building an agentic evaluation system for automatic scientific ideation rating.
This internship reflects the applied side of my work: turning model-centric research ideas into practical scientific-evaluation systems.
Scholarly reviewing and field participation
Reviewer: NeurIPS 2026, KDD 2026, ICDM 2025, KDD 2025
Service matters to me because it sharpens taste, exposes emerging directions early, and keeps me engaged with the standards by which the field evaluates new work.
Systems
OpenClaw is where I turn ideas about long-horizon reasoning, coordination, and memory into an actual system. The operating-system design is organized around three principles: departmentalized agents instead of one monolithic assistant, layered external memory instead of hidden context, and self-evolution through explicit write-back to skills, docs, and workspace artifacts. In practice, the system learns from user–agent interaction and from the outcomes of agent–agent collaboration, which makes it feel much closer to an in-context reinforcement-learning loop than to a static prompt wrapper. I include it here because it reflects a style of work I care about: turning research questions into concrete, inspectable systems.
My system is designed around specialized departments rather than a single generalist assistant. Research, coding, job-hunting, and other roles can be separated, coordinated, and supervised with explicit boundaries.
The memory system is deliberately externalized across daily notes, curated long-term memory, skills, docs, and workspace files so that context survives session resets and remains inspectable by both humans and agents.
The operating system learns from user–agent interaction and from the outcomes of agent–agent collaboration. By writing those lessons back into prompts, skills, docs, and workflows, my system behaves much closer to an in-context reinforcement-learning loop than to a static agent shell.
My long-term goal is to help build autonomous AI research systems that make serious research more widely accessible. High-quality research capacity is still concentrated in a relatively small number of institutions and teams. If AI systems can better use learned knowledge, reason more reliably, and interact productively with structured scientific evidence, they can broaden who gets to do meaningful research and at what scale.
Recent updates
Curiosity-Driven Questioning and InfRL extend the research-ideation line at KDD 2026.
Serving as Teaching Assistant for Artificial Intelligence (CS4710) at UVA.
Co-authored HyHG received the ICDM 2025 Best Paper Award.
Completed a Research Scientist internship at Autoscience Institute focused on automatic scientific ideation rating.
IdeaBench, InfAL, and hypothesis-generation work appeared at KDD, EMNLP, IJCAI, SDM, and Bioinformatics venues.