Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; its reference list is not available.

Evidence from counterfactual tasks supports emergent analogical reasoning in large language models

Open access

Abstract

A major debate has recently arisen concerning whether large language models (LLMs) have developed an emergent capacity for analogical reasoning. While some recent work has highlighted the strong zero-shot performance of these systems on a range of text-based analogy tasks, often rivaling human performance, other work has challenged these conclusions, citing evidence from so-called “counterfactual” tasks—tasks that are modified so as to decrease similarity with materials that may have been present in the language models’ training data. Here, we report evidence that language models are also capable of generalizing to these new counterfactual task variants when they are augmented with the ability to write and execute code. The results further corroborate the emergence of a capacity for analogical reasoning in LLMs and argue against claims that this capacity depends on simple mimicry of the training data.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1093/pnasnexus/pgaf135

Cites

Statistics

Figures

Fig. 1. Results for letter-string analogies with shuffled alphabet. A) Example of a counterfactual letter-string analogy problem constructed using a permuted alphabet. This example involves a successor transformation with an interval size of 1, applied to the final letter of the string. Other problems involved an interval size of 2, and different types of transformations. B) Summary of results for human participants, GPT-4, GPT-4 with zero-shot chain-of-thought, and a variant of GPT-4 augmented with the capacity to write and execute code. C) Results for problems with an interval size of 1, sorted by transformation type. D) Results for problems with an interval size of 2. Human results reflect average performance for N=99 participants for interval-size-1 and N=97 separate participants for interval-size-2. Human behavioral data were collected in an online experiment. The experiment was approved by the UCLA Institutional Review Board, and all participants provided informed consent. Black error bars represent standard error of the mean across participants. Gray error bars represent 95% binomial CIs for average model performance across multiple problems.
Fig. 1. Results for letter-string analogies with shuffled alphabet. A) Example of a counterfactual letter-string analogy problem constructed using a permuted alphabet. This example involves a successor transformation with an interval size of 1, applied to the final letter of the string. Other problems involved an interval size of 2, and different types of transformations. B) Summary of results for human participants, GPT-4, GPT-4 with zero-shot chain-of-thought, and a variant of GPT-4 augmented with the capacity to write and execute code. C) Results for problems with an interval size of 1, sorted by transformation type. D) Results for problems with an interval size of 2. Human results reflect average performance for N=99 participants for interval-size-1 and N=97 separate participants for interval-size-2. Human behavioral data were collected in an online experiment. The experiment was approved by the UCLA Institutional Review Board, and all participants provided informed consent. Black error bars represent standard error of the mean across participants. Gray error bars represent 95% binomial CIs for average model performance across multiple problems.CC BY-NC · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

No citation lists for this paper

No citation lists for this paper yet — checked again daily.

References 11 · Cited by 9

Metadata

Authors
Webb, T. W., Holyoak, K. J., & Lu, H.
Affiliations
  • Microsoft Research New York City (United States) (US)
  • University of California, Los Angeles (US)
Corresponding author
Webb, Taylor W
Venue
PNAS Nexus
Published
2025-04-30
Volume
4
Issue
5
Pages
pgaf135-pgaf135
ISSN
2752-6542 (ISSN-L)
OpenAlex
W4410781058
Cited by
9
References
11
Citations by year
2026: 4 · 2025: 5
FWCI
3.61
Type
journal article
Language
en
Open access
Locations
  • PNAS Nexus (publishedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
award from Microsoft Azure as part of the Accelerating Foundation Models Research initiative
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
Keywords
analogical reasoning, large language models, zero-shot generalization, code generation, program execution, emergent capabilities, analogy tasks

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Evidence from counterfactual tasks supports emergent analogical reasoning in large language models. https://app.cochat.ai/papers/10.1093/pnasnexus/pgaf135
The paper
Webb, T. W., Holyoak, K. J., & Lu, H. (2025). Evidence from counterfactual tasks supports emergent analogical reasoning in large language models. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgaf135

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.