İsmail İlkan Ceylan
Principal Investigator
Reasoning
How learning systems reason, from continuous reasoning in language models to neural model counting and logical query answering.
[TODO: review this text.]
Reasoning reliably over relational knowledge is a long-standing goal of AI. Our work connects learning with logic, from classic knowledge representation to the reasoning abilities of modern language models.
Key Directions
- Reasoning in language models: Continuous reasoning with multiplexed tokens, beyond step-by-step reasoning in natural language.
- Learning to reason: Neural networks for approximate model counting and weighted model integration.
- Logic and knowledge representation: Query answering and explanations over ontologies and probabilistic knowledge bases.
People
Selected Publications
All publications →MUX: Continuous Reasoning via Multiplexed Tokens
In NeurIPS 2026 (Spotlight)
December 2026
Thinking with Looped Flows
In arXiv preprint
September 2026
What are the Right Symmetries for Formal Theorem Proving?
In arXiv preprint
May 2026
One Model, Any Conjunctive Query: Graph Neural Networks for Answering Queries over Incomplete Knowledge Graphs
In Learning on Graphs Conference (LoG) 2025
December 2025
On the Approximability of Weighted Model Integration on DNF Structures
In KR 2020 (Best Student Paper Runner-Up)
September 2020
Learning to Reason: Leveraging Neural Networks for Approximate DNF Counting
In AAAI 2020
February 2020