A model can infer a pattern from examples and derive a conclusion from stated premises. Tom Zahavy’s 2026 position paper argues that those abilities do not, by themselves, explain how a system invents a new premise that makes sense of sparse observations. He calls that creative move a “jump.” It is a thesis about the limits of current language models, not a proven impossibility theorem or settled consensus.
What induction, deduction, and abduction mean
These terms describe different ways of reaching a conclusion. A simple illustration—not a result from any of the studies below—shows the distinction:
As an Amazon Associate I earn from qualifying purchases.
- Induction: You observe that several metal objects expand when heated and infer that metals generally expand when heated. The inference reaches beyond the examples, so it is plausible, not guaranteed by them alone.
- Deduction: You are told that all objects in a particular set expand when heated and that a certain object belongs to that set. You can validly conclude that the object expands when heated, assuming the premises are true.
- Abduction: You find that a metal rod has lengthened after being warmed and propose that heating explains the change. That is a candidate explanation; it must still be checked against observations and alternatives.
In Zahavy’s framing, language-model pattern learning is treated broadly as induction. That is a useful lens for the argument, not a complete account of how every model works. Deduction operates from premises: a valid derivation can show what follows from them, but cannot by itself establish that they describe the world.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallWhat Zahavy means by a model’s “jump”
Zahavy’s 2026 ICML position paper, “Position: LLMs can’t jump”, focuses on the move from observations or simulations to a new explanatory hypothesis or formal axiom. The proposed jump is not merely spotting a pattern or applying a rule. It is supplying a premise that could explain what has been observed, then reasoning from that premise.
#1 Best Overall
The paper uses Einstein’s formulation of general relativity as a computational case study. Zahavy argues that compressing data or generalizing from patterns does not fully account for scientific discovery, especially when observations are sparse. He identifies translating a simulation into formal axioms as a critical bottleneck in artificial scientific invention.
That is the paper’s argument, not a demonstrated proof that no current model can ever originate an explanatory idea. Zahavy’s proposed direction is to give models sensory grounding through physically consistent, multimodal world models. The paper presents that as a possible way to bridge the gap, not as an established solution.
Rank #2
What task-specific studies show about the limits
Experimental results help test broad claims about reasoning, but different tasks measure different things. Symbol manipulation, proposing rules, applying rules, and scientific discovery are not interchangeable capabilities.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Symbolic induction can break down as patterns grow
In a 2023 study, Jing Qian and colleagues tested language models on copy, reverse, and addition tasks. They reported that performance dropped quickly on certain tasks as the number of symbols or repetitions increased. Their tested methods before the tutor approach did not completely solve the simplest addition induction problem.
The authors’ tutor-based method reached 100% accuracy in the study’s specified out-of-domain and repeating-symbol situations. That figure describes those tested settings and that method; it is not a general accuracy claim about models or arithmetic. See the ACL Anthology paper.
Proposing a rule and using it are different skills
Linlu Qiu and colleagues studied hypothesis refinement: a model proposes candidate rules, a symbolic interpreter checks them against examples, and the model refines its candidates. They report that this hybrid process can produce strong results on several benchmarks, while also finding brittleness and gaps between generating rules and applying them. The study therefore shows how external checking can help evaluate candidate hypotheses; it does not establish that a model independently creates a scientific explanation. Read the 2023 preprint.
Training can improve out-of-domain performance in tested settings
A 2026 preprint by Mingzi Cao and colleagues reports that training on reasoning trajectories improved performance on its realistic out-of-domain tasks, with gains of up to 14.60 in that evaluation. The result is evidence of transfer under the study’s training and testing setup, not proof of unrestricted generalization or scientific invention. See the preprint abstract.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →How to read claims that AI can or cannot reason
A capability claim is easier to judge when you ask what kind of task was tested and what support the model received. The distinction matters because a result on a narrow symbolic benchmark cannot settle the broader question of scientific discovery.
Best Value
- What is the task? Is the model generalizing a pattern, deriving a consequence from given premises, proposing an explanatory hypothesis, or validating one?
- How familiar are the examples? Results on familiar examples, out-of-domain cases, and sparse observations address different kinds of transfer.
- What tools are involved? A tutor, symbolic interpreter, or other external system can change what the overall method accomplishes. Attribute a result to the full setup, not automatically to the model working alone.
- What kind of claim is being made? Zahavy offers a philosophical and conceptual position; the other papers report results on particular tasks and methods. A proposed research direction is not an experimental finding.
A further caution comes from the 2022 TMLR paper “Emergent abilities of large language models”. Its Google Research record describes abilities that were near random at smaller scales, making them difficult to predict by extrapolating a scaling law from those smaller models. This cautions against confident forecasts from limited measurements; it does not demonstrate an unbounded ability or resolve whether models can make abductive jumps.
What remains unresolved
Three questions are often collapsed into one: Can a system produce a novel hypothesis? Can it check that hypothesis against evidence? Does the result amount to a useful scientific explanation? A system might do one without doing all three. The studies discussed here illuminate particular pieces—symbolic pattern learning, rule refinement, and out-of-domain performance—but do not settle the larger question of whether language models can originate new scientific frameworks.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problems
