Uncertainty and forecasting
How can language models represent uncertainty in their knowledge, future outcomes, and other people's beliefs?
Examples: Forecasting outcomes in negotiations, on social media, and predicting when users are uncertain.
West Virginia University · CSEE
Anthony Sicilia
Assistant Professor
I study how large language models and other AI systems reason when information is uncertain, goals are unclear, or people are part of the loop. My work combines statistical tools for uncertainty with linguistic theories of ambiguity, context, and communication. In this view, language can become a computational interface between uncertain evidence and useful action.
News
- Fall 2026: NSF grant awarded (PI) on improving reliability in human-AI communication.
- Fall 2026: Top Reviewer at NeurIPS.
- Spring 2026: New papers accepted at EMNLP (Main), ACL (Main), ACM FAccT, and ICML.
Selected Research Directions
Collaboration and ambiguous goals
How should AI systems ask questions, repair misunderstandings, and refine goals before producing an answer or taking action?
Examples: Resolving ambiguous user goals in scientific coding workflows and forming collaborative plans with agentic AI systems.
Learning from imperfect data
How can statistical learning theory help us reason about bias, robustness, and effective use of human supervision when data are noisy or limited?
Examples: Achieving more fair human-AI conversations and learning from limited data in safety-critical scenarios.
Multimodal reasoning
How should AI systems reason when language connects to images, signed language, interfaces, instructions, and embodied context?
Examples: Building multimodal and accessible interfaces to AI systems.