"LLMs Can't Jump": the position paper arguing large language models have a fundamental reasoning ceiling

A position paper published at OpenReview challenges a core assumption of the AI scaling thesis: that reasoning capability improves indefinitely with scale. The argument isn't new—but the framing is sharper than usual.

The Fact A position paper titled “LLMs Can't Jump” was submitted to OpenReview (Aug 5, 2026, 30 HN points, 11 comments). Based on the title and Hacker News discussions—since the full paper is not yet publicly available—the thesis appears to argue that LLMs hit a fundamental ceiling on certain reasoning tasks: abilities that seem to emerge on simple examples collapse when faced with more complex variants, and scaling alone does not solve the problem.

Our Take If the thesis holds up under scrutiny, it aligns with a growing body of critical work—studies on the “reversal curse,” failures on ARC-AGI, analyses by François Chollet on generalization. The argument isn’t about raw performance (benchmarks keep improving) but about the nature of reasoning: pattern matching rather than compositional reasoning. The practical implication is that standard benchmarks (GSM8K, MATH) overestimate real capabilities because they lie within the training distribution. Treat this as provisional until full peer review: a title and HN score do not a verified thesis make.

Watch This Space The paper’s full release and responses from labs (OpenAI, Anthropic, DeepMind) will be telling, as they have a vested interest in refuting or qualifying these limitations if they prove accurate.