Tom Zahavy

Research scientist at Google DeepMind. Author of llms-cant-jump (January 2026), the position paper arguing that LLMs have mastered induction, are conquering deduction, and are structurally incapable of abduction.

The detail that gives the paper its weight: Zahavy is a co-author of the work he cites as the deduction success. He appears in the author list of Hubert et al. (2025) — alphaproof — and in Zahavy et al. (2024), the test-time-RL work that solved the Penrose chess position, which he cites in the same paper as an instance of inventing problem variations to force progress.

So the limitation argued here is argued from inside the programme, by someone who helped build the part that works, about the part that doesn’t. That is a different thing from an outside critic saying scaling won’t get there, and it is why the paper is the corpus’s strongest counter to “more compute closes any gap” — the person making it works at the lab with the most to gain from the opposite claim.

Related

llms-cant-jump · abduction · alphaproof · llms-cant-jump-press