pith. sign in

arxiv: 2606.13607 · v1 · pith:3QZT4OX7new · submitted 2026-06-11 · 💻 cs.AI

Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning

Zach Studdiford , Gary Lupyan This is my paper

Pith reviewed 2026-06-27 06:42 UTC · model grok-4.3

classification 💻 cs.AI
keywords pattern matchingLLM reasoninghuman reasoningattention headscommon-sense reasoningcausal reasoningerror patternseveryday reasoning
0
0 comments X

The pith

Everyday causal reasoning in humans and LLMs relies on pattern matching rather than abstract world models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests human participants and 25 LLMs on common-sense reasoning about everyday situations and finds matching patterns of errors in both. Specific attention heads in the LLMs drive these responses through pattern matching. Those heads predict human errors triggered by irrelevant details in the prompts. The results indicate that such reasoning in people and models follows from matching patterns instead of building abstract world models.

Core claim

We observe similar patterns of errors in both people and models. We identify the set of attention heads driving LLM responses and find that these heads implement a form of pattern-matching. These attention heads allow us to predict seemingly inexplicable reasoning errors in people caused by ostensibly irrelevant prompt details. Taken together, our results suggest that everyday causal reasoning in people and LLMs is more consistent with a form of pattern-matching than with abstract world models.

What carries the argument

Attention heads in LLMs that implement pattern-matching and predict human errors from irrelevant prompt details.

If this is right

  • Both humans and LLMs rely on pattern matching for everyday causal reasoning instead of abstract world models.
  • Attention heads identified in LLMs can forecast human errors on prompts with irrelevant details.
  • Error patterns observed in LLMs reflect mechanisms also present in human reasoning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same pattern-matching approach might apply to other reasoning domains beyond causal everyday tasks.
  • Efforts to instill abstract world models in LLMs may not reduce errors if the underlying process remains pattern-based.
  • Altering the relevant attention heads could produce targeted changes in both model and predicted human behavior on new prompts.

Load-bearing premise

The assumption that the identified attention heads implement the same pattern-matching process that produces the parallel error patterns in human participants, shown by their success at predicting those errors.

What would settle it

A new set of everyday reasoning tasks where the identified LLM attention heads fail to predict the specific errors made by human participants.

Figures

Figures reproduced from arXiv: 2606.13607 by Gary Lupyan, Zach Studdiford.

Figure 1
Figure 1. Figure 1: Overview of our evaluation probing human and LLM causal reasoning. a. Illustration of the evaluation format used for testing causal reasoning in people and LLMs. For each prompt, subjects are first presented with the scenario. After reading the prompt, subjects press SPACE to see the two response options and then select the most likely completion. b. Summary of the 11 categories we used in our evaluation o… view at source ↗
Figure 2
Figure 2. Figure 2: Behavioral results on our reasoning task ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of inconsistencies in human and LLM reasoning. Top: minimal counterfactual changes to causal reasoning problems (changing one word to modify the ground-truth prompt outcome) elicit similar fail states in people and models. For each prompt pair, the left axis tick indicates standardized human and gemma-3-27b accuracy for the lower prompt, while the right axis tick shows human and gemma-3-27b for th… view at source ↗
Figure 4
Figure 4. Figure 4: Identifying the internal mechanisms underlying LLM responses (results included here are for [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Predicting human accuracy from causally important attention heads. a. Ground-truth human accuracy (y axis) versus leave-one-out predictions of human accuracy from a linear model with the top-k attention heads and fixed effects for prompt category. The dotted gray line indicates perfect fit (y = yˆ); solid lines show best fits for each category. b. Comparisons of R2 distributions from linear model permutati… view at source ↗
Figure 6
Figure 6. Figure 6: Content effects in human reasoning are explained by causally important attention heads. a.Human accuracies for six sets of minimally different scenarios. Labels above show scenario templates, words on x axis ticks indicate words substituted in the [obj] template slot for each scenario. b. ground truth human accuracy versus held out model predictions from a regression with top-k attention heads (k = 7 selec… view at source ↗
Figure 7
Figure 7. Figure 7: Human evaluation trial sequence. Participants are first presented with the prompt and asked to press SP ACE to reveal answer options. They are then instructed to press z or / corresponding to an answer choice (order randomized). state_nonfollows two_refs_egocentric two_refs_geocentric geocentric_distant geocentric_near relative_pos state_follows action_follows action_nonfollows egocentric_distant egocentri… view at source ↗
Figure 8
Figure 8. Figure 8: Consistency between first and second responses for a given subject, separated by category. X-axis [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Mean accuracy of open and closed source models as given by accuracy measured from model [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Mean accuracy of open and closed source models as given by the difference of correct and incorrect [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: R2 values for linear models predicting human accuracy from LLM generation accuracy. *** *** *** *** *** *** * 0.0 0.1 0.2 0.3 gemma−3−27b−it gemma−2−27b−it gemma−2−9b−it gemma−3−12b−it gemma−2−27b gemma−3−4b−it llama−2−7b−chat gemma−2−2b−it llama−3.1−8b gemma−2−2b gpt2−xl llama−2−13b gemma−2−9b llama−2−13b−chat llama−2−7b gpt2−large llama−3.1−8b−instruct gemma−3−27b gemma−3−4b gpt2−medium gemma−3−12b R 2 … view at source ↗
Figure 12
Figure 12. Figure 12: R2 values for linear models predicting human accuracy from LLM logit differences. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Human accuracy versus logit difference for Gemma 3 12B IT Category-level averages showing the relationship between model logit difference and human accuracy, with dotted lines connecting context and target prompt types within each category. action_follows action_follows action_nonfollows action_nonfollows egocentric_distant egocentric_distant egocentric_near egocentric_near geocentric_distant geocentric_d… view at source ↗
Figure 14
Figure 14. Figure 14: Human accuracy versus logit difference for Gemma 2 27B IT Category-level averages showing the relationship between model logit difference and human accuracy, with dotted lines connecting context and target prompt types within each category. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Human accuracy versus logit difference for Gemma 2 9B IT Category-level averages showing the relationship between model logit difference and human accuracy, with dotted lines connecting context and target prompt types within each category. action_follows action_follows action_nonfollows action_nonfollows egocentric_distant egocentric_distant egocentric_near egocentric_near geocentric_distant geocentric_di… view at source ↗
Figure 16
Figure 16. Figure 16: Human accuracy versus logit difference for Gemma 3 4B IT Category-level averages showing the relationship between model logit difference and human accuracy, with dotted lines connecting context and target prompt types within each category. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Human accuracy versus logit difference for Gemma 2 2B IT Category-level averages showing the relationship between model logit difference and human accuracy, with dotted lines connecting context and target prompt types within each category. A.6.1 Proprietary model category alignment For proprietary models, we measure accuracy using binary scores of token outputs. We include results below for all evaluated … view at source ↗
Figure 18
Figure 18. Figure 18: Human accuracy versus model correctness for DeepSeek R1 Category-level averages showing the relationship between model accuracy and human accuracy, with dotted lines connecting context and target prompt types within each category. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Human accuracy versus model correctness for GPT-5.2 Category-level averages showing the relationship between model accuracy and human accuracy, with dotted lines connecting context and target prompt types within each category. action_follows action_follows action_nonfollows action_nonfollows egocentric_distant egocentric_distant egocentric_near egocentric_near geocentric_distant geocentric_distant geocent… view at source ↗
Figure 20
Figure 20. Figure 20: Human accuracy versus model correctness for GPT-4.1 Category-level averages showing the relationship between model accuracy and human accuracy, with dotted lines connecting context and target prompt types within each category. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Human accuracy versus model correctness for Gemini 2.0 Category-level averages showing the relationship between model accuracy and human accuracy, with dotted lines connecting context and target prompt types within each category. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Item-Level Human accuracy versus logit difference for Gemma 3 12B IT Human accuracy plotted against model logit difference at the item level, faceted by category and colored by prompt type (context vs target). 29 [PITH_FULL_IMAGE:figures/full_fig_p029_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Item-Level Human accuracy versus logit difference for Gemma 2 27B IT Human accuracy plotted against model logit difference at the item level, faceted by category and colored by prompt type (context vs target). state_nonfollows two_refs_egocentric two_refs_geocentric geocentric_distant geocentric_near relative_pos state_follows action_follows action_nonfollows egocentric_distant egocentric_near −10 0 10 20… view at source ↗
Figure 24
Figure 24. Figure 24: Item-Level Human accuracy versus logit difference for Gemma 2 9B IT Human accuracy plotted against model logit difference at the item level, faceted by category and colored by prompt type (context vs target). 30 [PITH_FULL_IMAGE:figures/full_fig_p030_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Item-Level Human accuracy versus logit difference for Gemma 3 4B IT Human accuracy plotted against model logit difference at the item level, faceted by category and colored by prompt type (context vs target). state_nonfollows two_refs_egocentric two_refs_geocentric geocentric_distant geocentric_near relative_pos state_follows action_follows action_nonfollows egocentric_distant egocentric_near −10 0 10 −10… view at source ↗
Figure 26
Figure 26. Figure 26: Item-Level Human accuracy versus logit difference for Gemma 2 2B IT Human accuracy plotted against model logit difference at the item level, faceted by category and colored by prompt type (context vs target). 31 [PITH_FULL_IMAGE:figures/full_fig_p031_26.png] view at source ↗
Figure 27
Figure 27. Figure 27 [PITH_FULL_IMAGE:figures/full_fig_p032_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Attention head ablation results for gemma-2-27b-it. Heatmap values indicate the average change in logit difference after zero-ablating the activations of a given head for all prompts in the causal reasoning evaluation. 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Head Layer |Avg Impact| 0.1 0.2 0.3… view at source ↗
Figure 29
Figure 29. Figure 29: Attention head ablation results for gemma-2-9b-it. Heatmap values indicate the average change in logit difference after zero-ablating the activations of a given head for all prompts in the causal reasoning evaluation. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Model R2 and BIC at different k. We fit linear models using the entropy values for the top-k attention heads [PITH_FULL_IMAGE:figures/full_fig_p040_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Per-scenario predictions of human acccuracy. Y-axis indicates ground-truth mean human accuracy. X-axis shows predicted human accuracy from the top seven attention head activations. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Model R2 and BIC at different k. We refit linear models using the entropy values for the top-k attention heads [PITH_FULL_IMAGE:figures/full_fig_p045_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Per-scenario predictions of human acccuracy. Y-axis indicates ground-truth mean human accuracy. X-axis shows predicted human accuracy from the top seven attention head activations. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: Predicting human accuracy for left-out scenarios. Each X-axis coordinate indicates predictions for a prompt where that prompt’s scenario was not included in fitting regression model weights. 47 [PITH_FULL_IMAGE:figures/full_fig_p047_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: Comparing attention head predictiveness of human accuracy versus causal importance to model [PITH_FULL_IMAGE:figures/full_fig_p048_35.png] view at source ↗
read the original abstract

When large language models (LLMs) fail to generalize or make haphazard errors in reasoning, it is often taken as evidence that LLMs are not truly reasoning, but rather performing a kind of pattern matching. The implication is that people's behavior does not exhibit the same types of failures because human reasoning uses principled and abstract world models. We evaluate human participants and 25 LLMs on their ability to engage in common-sense reasoning about a variety of everyday situations and observe similar patterns of errors in both people and models. We then identify the set of attention heads driving LLM responses and find that these heads implement a form of pattern-matching. These attention heads allow us to predict seemingly inexplicable reasoning errors in people caused by ostensibly irrelevant prompt details. Taken together, our results suggest that everyday causal reasoning in people and LLMs is more consistent with a form of pattern-matching than with abstract world models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The paper claims that human participants and 25 LLMs exhibit similar error patterns (influence of irrelevant details, haphazard failures) on everyday causal reasoning tasks; that specific attention heads in the LLMs implement pattern-matching; and that these heads predict human errors from ostensibly irrelevant prompt details. On this basis the authors conclude that everyday causal reasoning in both people and LLMs is more consistent with pattern-matching than with abstract world models.

Significance. If the parallel error patterns and the predictive link from LLM attention heads to human errors survive rigorous controls and statistical evaluation, the work would supply a concrete mechanistic bridge between human and model behavior and would weaken the common claim that human reasoning relies on abstract models while LLMs rely only on surface patterns. The approach of using identified LLM components to generate falsifiable predictions about human performance is a methodological strength.

major comments (3)
  1. [Abstract] Abstract: the inference that the observed error signatures are 'more consistent with a form of pattern-matching than with abstract world models' is load-bearing for the central claim yet rests on an unexamined assumption. The manuscript provides no argument or data showing that an agent possessing an abstract causal model could not produce the same surface errors under incomplete retrieval or prompt interference; therefore the incompatibility conclusion does not follow from the reported parallels alone.
  2. [Abstract] Abstract: no experimental details, participant numbers, task descriptions, statistical tests, or controls are supplied for either the human or LLM evaluations. Without these it is impossible to determine whether the reported error patterns are reliable, replicable, or actually support the shared-mechanism conclusion.
  3. [Abstract] Abstract: the claim that the identified attention heads 'implement a form of pattern-matching' and 'allow us to predict' human errors is central to the mechanistic argument, yet the manuscript gives no description of how the heads were isolated, what their activation patterns are, or how the prediction to human data was performed and validated.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We agree that it requires expansion for experimental details and methodological descriptions to allow proper evaluation. We also acknowledge the need to clarify the logical basis for our conclusion of greater consistency with pattern-matching. We address each major comment below and will make corresponding revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the inference that the observed error signatures are 'more consistent with a form of pattern-matching than with abstract world models' is load-bearing for the central claim yet rests on an unexamined assumption. The manuscript provides no argument or data showing that an agent possessing an abstract causal model could not produce the same surface errors under incomplete retrieval or prompt interference; therefore the incompatibility conclusion does not follow from the reported parallels alone.

    Authors: We thank the referee for highlighting this. Our claim is not one of strict incompatibility but of greater consistency, supported by the fact that the specific pattern-matching attention heads identified in LLMs predict human error rates on ostensibly irrelevant details. This predictive link would be surprising if human reasoning relied exclusively on abstract causal models that retrieve only relevant structures. That said, we agree an explicit argument addressing why abstract models should not exhibit these errors under the task conditions would strengthen the manuscript. We will add a paragraph in the Discussion section making this case. revision: partial

  2. Referee: [Abstract] Abstract: no experimental details, participant numbers, task descriptions, statistical tests, or controls are supplied for either the human or LLM evaluations. Without these it is impossible to determine whether the reported error patterns are reliable, replicable, or actually support the shared-mechanism conclusion.

    Authors: We agree that the abstract omits these elements, which are provided in the full manuscript's Methods and Results sections. In the revised version we will expand the abstract to include concise summaries of participant numbers, task descriptions, statistical tests, and controls so that the reliability of the error patterns can be assessed from the abstract alone. revision: yes

  3. Referee: [Abstract] Abstract: the claim that the identified attention heads 'implement a form of pattern-matching' and 'allow us to predict' human errors is central to the mechanistic argument, yet the manuscript gives no description of how the heads were isolated, what their activation patterns are, or how the prediction to human data was performed and validated.

    Authors: The abstract is intentionally high-level, but we accept that it should convey more about the approach. We will revise it to briefly note the isolation method (attention analysis combined with causal interventions), the relevant activation patterns, and the validation of the prediction to human data, drawing directly from the detailed description already present in the Methods and Results sections of the manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: central claims rest on independent empirical comparisons of error patterns and attention-head interventions.

full rationale

The derivation proceeds from collecting human and LLM responses on everyday reasoning tasks, observing parallel error patterns, identifying attention heads in LLMs via standard interpretability methods, and testing whether those heads predict human errors from irrelevant prompt details. None of these steps reduces by definition or by self-citation to the target conclusion; the pattern-matching interpretation is an inference from the data rather than presupposed in the measurement. No equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided abstract or described methods. The work is therefore self-contained against external benchmarks of human and model behavior.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical derivations, fitted parameters, or invented entities are described in the abstract.

pith-pipeline@v0.9.1-grok · 5674 in / 969 out tokens · 24091 ms · 2026-06-27T06:42:28.872822+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 8 canonical work pages

  1. [1]

    Renee Baillargeon

    URLhttp://arxiv.org/abs/2308.03762. Renee Baillargeon. Physical reasoning in infancy.The cognitive neurosciences, pages 181–204,

    arXiv

  2. [2]

    Characterizing pattern matching and its limits on compositional task structures.arXiv preprint arXiv:2505.20278,

    Hoyeon Chang, Jinho Park, Hanseul Cho, Sohee Yang, Miyoung Ko, Hyeonbin Hwang, Seungpil Won, Dohaeng Lee, Youbin Ahn, and Minjoon Seo. Characterizing pattern matching and its limits on compositional task structures.arXiv preprint arXiv:2505.20278,

    arXiv

  3. [3]

    Llms are not (consistently) bayesian: Quantifying internal (in) consistencies of llms’ probabilistic beliefs.arXiv preprint arXiv:2605.06915,

    Chacha Chen, Matthew Jörke, Adam Goliński, Masha Fedzechkina, Guillermo Sapiro, Sinead Williamson, and Nicholas Foti. Llms are not (consistently) bayesian: Quantifying internal (in) consistencies of llms’ probabilistic beliefs.arXiv preprint arXiv:2605.06915,

    Pith/arXiv arXiv

  4. [4]

    A systematic comparison of syllogistic reasoning in humans and language models

    Tiwalayo Eisape, Michael Tessler, Ishita Dasgupta, Fei Sha, Sjoerd Steenkiste, and Tal Linzen. A systematic comparison of syllogistic reasoning in humans and language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 8425–8444,

    2024

  5. [5]

    doi: 10.3758/BF03196976

    ISSN 1532-5946. doi: 10.3758/BF03196976. URLhttps://doi.org/10.3758/BF03196976. Chaz Firestone. Performance vs. competence in human–machine comparisons.Proceedings of the National Academy of Sciences, 117(43):26562–26571,

    work page doi:10.3758/bf03196976

  6. [6]

    A theory of emergent in-context learning as implicit structure induction

    Michael Hahn and Navin Goyal. A theory of emergent in-context learning as implicit structure induction. arXiv preprint arXiv:2303.07971,

    arXiv

  7. [7]

    Linear latent world models in simple transformers: A case study on othello-gpt.arXiv preprint arXiv:2310.07582,

    60 Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning Dean S Hazineh, Zechen Zhang, and Jeffery Chiu. Linear latent world models in simple transformers: A case study on othello-gpt.arXiv preprint arXiv:2310.07582,

    arXiv

  8. [8]

    In-context learning creates task vectors.arXiv preprint arXiv:2310.15916,

    Roee Hendel, Mor Geva, and Amir Globerson. In-context learning creates task vectors.arXiv preprint arXiv:2310.15916,

    arXiv

  9. [9]

    Prompting is not a substitute for probability measurements in large language models.arXiv preprint arXiv:2305.13264,

    Jennifer Hu and Roger Levy. Prompting is not a substitute for probability measurements in large language models.arXiv preprint arXiv:2305.13264,

    arXiv

  10. [10]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    Pith/arXiv arXiv 2001

  11. [11]

    How do language models compose functions?arXiv preprint arXiv:2510.01685,

    Apoorv Khandelwal and Ellie Pavlick. How do language models compose functions?arXiv preprint arXiv:2510.01685,

    Pith/arXiv arXiv

  12. [12]

    doi: 10.1016/0010-0285(85)90009-X

    ISSN 00100285. doi: 10.1016/0010-0285(85)90009-X. URLhttps://linkinghub.elsevier.com/retrieve/pii/001002858590009X. John K Kruschke and Michael M Fragassi. The perception of causality: Feature binding in interacting objects. InProceedings of the eighteenth annual conference of the cognitive science society, pages 441–446. Routledge,

    work page doi:10.1016/0010-0285(85)90009-x

  13. [13]

    Language models struggle to use representations learned in-context.arXiv preprint arXiv:2602.04212,

    61 Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning Michael A Lepori, Tal Linzen, Ann Yuan, and Katja Filippova. Language models struggle to use representations learned in-context.arXiv preprint arXiv:2602.04212,

    Pith/arXiv arXiv

  14. [14]

    Using Counterfactual Tasks to Evaluate the Generality of Analogical Reasoning in Large Language Models

    Martha Lewis and Melanie Mitchell. Using Counterfactual Tasks to Evaluate the Generality of Analogical Reasoning in Large Language Models. URLhttp://arxiv.org/abs/2402.08955. Falk Lieder and Thomas L. Griffiths. Resource-rational analysis: Understanding human cognition as the optimal use of limited computational resources.The Behavioral and Brain Sciences...

    arXiv

  15. [15]

    Resource-Rational Analysis: Understanding Human Cognition as the Optimal Use of Limited Computational Resources.Behavioral and Brain Sciences2020,43, e1

    ISSN 1469-1825. doi: 10.1017/S0140525X1900061X. Tania Lombrozo. The structure and function of explanations.Trends in cognitive sciences, 10(10):464–470,

    work page doi:10.1017/s0140525x1900061x

  16. [16]

    doi: 10.1016/j.cognition.2013.08.015

    ISSN 0010-0277. doi: 10.1016/j.cognition.2013.08.015. URLhttp://www.sciencedirect. com/science/article/pii/S0010027713001728. Kyle Mahowald, Anna A Ivanova, Idan A Blank, Nancy Kanwisher, Joshua B Tenenbaum, and Evelina Fedorenko. Dissociating language and thought in large language models.Trends in cognitive sciences, 28 (6):517–540,

    work page doi:10.1016/j.cognition.2013.08.015 2013

  17. [17]

    Embers of autoregression: Understanding large language models through the problem they are trained to solve.arXiv preprint arXiv:2309.13638,

    R Thomas McCoy, Shunyu Yao, Dan Friedman, Matthew Hardy, and Thomas L Griffiths. Embers of autoregression: Understanding large language models through the problem they are trained to solve.arXiv preprint arXiv:2309.13638,

    arXiv

  18. [18]

    James A Michaelov, Seana Coulson, and Benjamin K Bergen

    ISBN 978-0-674-36830-9. James A Michaelov, Seana Coulson, and Benjamin K Bergen. Can peanuts fall in love with distributional semantics?arXiv preprint arXiv:2301.08731,

    arXiv

  19. [19]

    URLhttps://doi.org/10.1037/14043-011

    doi: 10.1037/14043-011. URLhttps://doi.org/10.1037/14043-011. Kanishka Misra and Kyle Mahowald. Language models learn rare phenomena from less rare phenomena: The case of the missing aanns. InProceedings of the 2024 conference on empirical methods in natural language processing, pages 913–929,

    work page doi:10.1037/14043-011 2024

  20. [20]

    Krakauer , title =

    doi: 10.1073/pnas.2215907120. URLhttps://www.pnas.org/doi/10.1073/pnas. 2215907120. Neel Nanda and Joseph Bloom. Transformerlens. https://github.com/TransformerLensOrg/ TransformerLens,

    work page doi:10.1073/pnas.2215907120

  21. [21]

    In-context learning and induction heads.arXiv preprint arXiv:2209.11895,

    Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. In-context learning and induction heads.arXiv preprint arXiv:2209.11895,

    Pith/arXiv arXiv

  22. [22]

    doi: 10.1017/S0140525X22002849

    ISSN 1469-1825. doi: 10.1017/S0140525X22002849. D.E. Rumelhart, J.L. McClelland, and the PDP Research Group.Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Volumes 1 and

    work page doi:10.1017/s0140525x22002849

  23. [23]

    Conceptual structure coheres in human cognition but not in large language models

    Siddharth Suresh, Kushin Mukherjee, Xizheng Yu, Wei-Chun Huang, Lisa Padua, and Timothy Rogers. Conceptual structure coheres in human cognition but not in large language models. InProceedings of the 2023 conference on empirical methods in natural language processing, pages 722–738,

    2023

  24. [24]

    Toward a theory of generalizability in llm mechanistic interpretability research.arXiv preprint arXiv:2509.22831,

    Sean Trott. Toward a theory of generalizability in llm mechanistic interpretability research.arXiv preprint arXiv:2509.22831,

    arXiv

  25. [25]

    Large language models fail on trivial alterations to theory-of-mind tasks.arXiv preprint arXiv:2302.08399,

    Tomer Ullman. Large language models fail on trivial alterations to theory-of-mind tasks.arXiv preprint arXiv:2302.08399,

    arXiv

  26. [26]

    doi: 10.1016/j.tics.2024.02.008

    ISSN 1364-6613. doi: 10.1016/j.tics.2024.02.008. URL https://www.sciencedirect.com/science/article/pii/S1364661324000354. Lance Ying, Aydan Y Huang, Aviv Netanyahu, Andrei Barbu, Boris Katz, Joshua B Tenenbaum, and Tianmin Shu. Grounding social perception in intuitive physics.arXiv preprint arXiv:2603.27410,

    work page doi:10.1016/j.tics.2024.02.008 2024