30 November 2026, 11:15–12:00 CET · IPI PAN, Jana Kazimierza 5, Warsaw, and online
I’ll be speaking at AI without Borders, the second day of the MOST 2026 mini-conference organized by the Linguistic Engineering Group at the Institute of Computer Science, Polish Academy of Sciences. The day brings together Polish AI researchers working around the world.
Large language models now produce reasoning that looks like proof, but have they learned the rules of inference or just the patterns that usually come with them? In joint work with Manuel Vargas Guzmán and Maciej Malicki, we use syllogistic logic as a controlled testbed and ask models which premises are needed to derive a conclusion. This separates two abilities usually lumped together as “compositional generalization”: extending a learned pattern to longer inferences, and recovering the underlying rules to apply them to simpler cases. Models are much better at the first than the second. I’ll also show that task-specific training substantially helps smaller models, and that even imperfect neural reasoners make useful assistants to a symbolic prover.
The event is hybrid and free to attend. Program and registration: most-pan-2026.github.io
