ICML Position Paper: Models Can Deduce, But They Cannot Jump

Tom Zahavy’s ICML 2026 position paper argues that language models have mastered induction, are rapidly conquering deduction, and lack any mechanism for abduction — the generation of new explanatory premises. It opens with Einstein’s letter to Maurice Solovine describing discovery as a non-logical jump to a fresh set of axioms, from which consequences are only afterwards deduced and tested. The claim is that without sensory grounding, that axiom-generating step cannot be performed. Zahavy has publicly stressed this is a personal position, not his employer’s view and not a claim that models can never discover anything. The practical reading: scope agents to execution and verification; keep humans on premise selection.