LLMs Can't Jump and They Shouldn't Have To

Tom Zahavy of Google DeepMind released a position paper titled “LLMs can’t jump” on OpenReview, contending that current generative models are strong at induction—scouring massive corpora for correlations—and increasingly capable of deduction through tools like AlphaProof, yet they stumble on abduction, the creative step of proposing unseen causes. Zahavy attributes this gap to the models’ lack of sensorimotor experience; they have never felt gravity or lifted a weight, so they cannot generate the kind of intuitive hypotheses that drove breakthroughs such as germ theory or Einstein’s relativity. The paper reframes this limitation as a design specification: instead of forcing LLMs to generate novel theories, developers should focus on their ability to ingest and synthesize cross‑disciplinary literature, surface contradictions, and verify formal arguments, leaving the abductive leap to human scientists.

The argument aligns with a broader industry trend of decentralizing AI capability. Meta’s recent push for open‑weight models exemplifies a strategy that favors distributing smaller, versatile models to many researchers rather than concentrating compute in a single, ever‑larger system. This mirrors the “centaur” model in chess, where human intuition and engine calculation combine to outperform either alone. By positioning LLMs as the “engine” that handles exhaustive reading and logical checking, the DeepMind paper suggests a practical workflow: the model curates a well‑posed problem, the researcher supplies the creative hypothesis, and a verification system like AlphaProof tests it. This division of labor could mitigate the diminishing returns of scaling compute, which have become contentious as hardware costs and energy consumption rise.

If the community adopts this collaborative paradigm, the bottleneck in scientific progress may shift from idea generation to effective integration of AI‑driven synthesis tools into everyday research pipelines. Risks include over‑reliance on model‑generated literature maps that might miss nuanced context, and the possibility that the “abduction” step remains inaccessible to those without deep domain expertise, potentially widening gaps between well‑funded labs and smaller teams. Watching how open‑weight model ecosystems evolve, how verification tools scale, and whether funding bodies incentivize human‑AI co‑discovery will be critical to gauge the real impact of this approach.