A position paper on OpenReview, titled «LLMs Can’t Jump», reaches 165 points on Hacker News with 116 comments. The thesis in the title: language models have a fundamental limit in making logical leaps, that is, in connecting information that is not close in the space of what they have seen during training.

Why it matters to you. If you build workflows that expect from AI the insight that connects two distant ideas, this paper says the insight doesn’t come on its own. Models interpolate well within their training data, but struggle to extrapolate beyond it. The difference matters when the task requires a step that isn’t already in the material seen: there the AI returns the closest plausible answer; the right answer requires a logical leap.

It’s the same thread we followed when discussing how the expertise of whoever guides the model matters more than prompting tricks: the person who knows the domain makes the logical leap, the AI articulates and develops it. A paper that formalizes this limit tells you where to delegate and where to keep human judgment in the loop.

In detail

The OpenReview page of the paper was not accessible at the time of collection (requires browser verification), so what follows is based on the title, the context of discussion on Hacker News, and the research thread the site has followed in recent weeks.

What are position papers. In machine learning venues like OpenReview, a “position paper” does not present a new experiment: it argues a point of view, often to stimulate debate on a research direction or on a limitation the community underestimates. The format allows you to say “let’s stop and think” without necessarily showing accuracy curves.

The leap that’s missing. The title uses a sports metaphor: the horse that won’t jump the hurdle. In the context of reasoning, a “logical leap” is that passage that connects two ideas or two pieces of data that are not close in the space of information seen during training. Models are good at interpolating: finding the answer that sits between two things they know. They struggle more with extrapolating: taking the step that goes beyond what they have seen.

This is not a new limitation in itself. Literature on LLM reasoning has long discussed the difference between interpolation and extrapolation. The paper seems to argue that this limitation is structural, not a bug to fix with more data or more parameters. The point matters because, if true, it changes expectations: a larger model answers questions better that resemble its training data, but doesn’t necessarily make the creative connection that a human expert sees.

Where it fits in the site’s thread. On August 4 we covered how Sean Goedecke’s analysis showed that the expertise of whoever guides the model matters more than prompting tricks: real domain competence produces better output because whoever has it knows where to find the logical leap. The same day, the paper on GradCuit showed that optimizing internal states at test time improves reasoning without more tokens. Two different angles on the same problem: how much of reasoning can we delegate and how much remains irreducibly human.

What we don’t know yet. Without access to the full text, we cannot verify whether the paper presents concrete examples of failed logical leaps, whether it proposes metrics to measure them, or whether it limits itself to a theoretical argument. The 116 comments on Hacker News suggest the thesis has sparked debate, but we cannot cite details of objections without having read them. The paper should be read before drawing operational conclusions.