Position: LLMs can't jump

Summary

ICML 2026 position paper. Einstein’s Solovine letter as a computational case study: induction (pattern matching) and deduction (proof from axioms) do not explain discoveries made from scarce observation. Abduction — inventing a cause/axiom via embodied simulation — is the missing mechanism; LLMs can plausibly run the deductive phase but are structurally Chinese-Room-like on the jump. The bottleneck is translating simulation into formal axioms. Proposed path: physically consistent, action-controllable multimodal world models rather than next-frame video generators. Discord posted a follow-up X article by Shubham Mishra arguing JEPA-style world models (I-JEPA, VL-JEPA, Ha & Schmidhuber 2018, DreamerV3) fill that gap; I-JEPA is already papers_local 590.

Keywords

abduction · world-models · jepa · scientific-discovery · deepmind · icml-2026 · x

Topics

scientific discovery, abduction, world models

Research notes

• Primary: project page / ICML 2026 position paper. PDF https://www.tomzahavy.com/files/llms-cant-jump.pdf and https://openreview.net/pdf?id=klU4737opt. Discord/X https://x.com/shubh6200/status/2084470616976187833 via fxtwitter (quotes his earlier recap https://x.com/shubh6200/status/2084338358214594772). Cataloged the Zahavy paper as the follow-through; I-JEPA already papers_local 590. Position paper, not a hosted corpus, so no datasets_local row.