Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

The Other Mind: How Language Models Exhibit Human Temporal Cognition

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Large language models, at sufficient scale, organize their judgments of year similarity around an internal temporal reference point and compress perceived distance from that point logarithmically.

desk verdict Multi-level temporal-cognition in LLMs is a fresh package, but the fixed reference point in dref is the center of the stimulus range, so the spontaneous-reference-point claim needs a stronger model comparison before it holds. read the letter →

arxiv 2507.15851 v1 pith:O36PK7SZ submitted 2025-07-21 cs.AI

Lingyu Li , Yang Yao , Yixu Wang , Chubo Li , Yan Teng , Yingchun Wang This is my paper
classification cs.AI
keywords temporalcognitionsimilarityjudgmentWeber-Fechnerlawsubjectivereferencepointlargelanguagemodelsmechanisticinterpretabilitylinearprobestemporal-preferentialneurons
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether large language models, without any explicit instruction about time, build a sense of time similar to our own. Using a similarity judgment task over every pair of years from 1525 to 2524, it claims that larger models spontaneously anchor their year judgments at a subjective reference point near 2025 and compress perceived distance from that point logarithmically, matching the Weber-Fechner law of human psychophysics. The finding matters because it would mean human-like cognitive structure can emerge from statistical prediction over text, rather than from a hand-built temporal module. The paper then traces the same pattern into the model's neurons, its layer-by-layer representations, and the statistical geometry of the pretraining corpus.

What carries the argument

The load-bearing construct is the reference-log-linear distance $d_{\mathrm{ref}}(i,j)=|\log(|R-i|)\circ\log(|R-j|)|$, a metric that expresses Weber-Fechner compression around a fixed subjective present: years on the same side of $R$ are compared by subtracting their log-distances, years on opposite sides by adding them. The paper pairs this predictor with two established tools: temporal-preferential neurons, selected by Cohen's $d$, FDR-corrected $t$-tests, and consistency across years; and linear probes, trained separately at each layer to decode how much of each candidate distance is linearly readable from hidden states. A third tool, multidimensional scaling of embeddings from pretrained encoders, exposes the temporal structure of the training corpus itself.

What would settle it

Recompute the behavioral regressions with each model's own estimated reference year (for example, the sliding-window estimates the paper already reports) instead of the fixed $R=2025$, and check whether the neuron activation trough lines up with that estimate. If no model's own-reference fit matches or beats the fixed-2025 fit, or if the activation minimum does not track the estimated reference, the claim that models spontaneously establish a subjective temporal anchor would be refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that temporal cognition in LLMs is organized by a subjective 'now', not by objective calendar position. In the year-to-year similarity task, larger models produce matrices in which years near a reference point are sharply separated while distant years collapse together; judged dissimilarity is best predicted by the reference-log-linear distance $d_{\mathrm{ref}}(i,j)=|\log(|R-i|)\circ\log(|R-j|)|$ with $R$ fixed at 2025, rather than by raw log-linear number distance or by string-edit distance. Three convergent signatures support the behavioral finding: a small subpopulation of temporal-preferential neurons whose activation is minimal at the reference year and grows logarithmically away from it; linear probes showing that years begin as numerical values in early layers and become reference-centered temporal orientations in deep layers; and pretrained embedding models whose semantic year geometry already contains this nonlinear, reference-centered structure. The authors interpret the whole pattern as evidence for an experientialist view in which the model actively constructs a subjective world model from its architecture and data.

Load-bearing premise

The load-bearing premise is that a single fixed reference point, $R=2025$, can stand for every model's subjective present when fitting the reference-log-linear distance; if a given model's true internal anchor is elsewhere, the main regressions do not actually test whether that model spontaneously established its own temporal reference point.

Editorial extensions

If this is right

  • In larger models, the best predictor of year-pair similarity is distance from the model's own reference point, not raw numerical difference, so temporal orientation can be measured behaviorally without introspection.
  • Temporal-preferential neurons implement a logarithmic code around the reference year, suggesting the Weber-Fechner compression is not prompt mimicry but an internal activation pattern.
  • Year representations are built hierarchically: numerical coding appears early, reference-centered temporal coding appears later, and in some models the late emergence of temporal coding suppresses the earlier numerical code.
  • The pretraining corpus already carries a nonlinear temporal geometry, so part of the apparent temporal cognition may be inherited from data statistics rather than invented by the model.
  • If these claims hold, model scale is what unlocks this temporal orientation: the smallest tested models show none of these patterns.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit a clean test: if the reference point is genuinely subjective, telling a model in the prompt that today is a far-future year should move the behavioral and neural reference point; this is a direct experimental follow-up.
  • The main regressions fix $R=2025$ for all models even though the paper's own sliding-window estimates range from 2010 to 2024; refitting with each model's estimated origin is the natural robustness check and could change which models look temporally organized.
  • The future-years-are-closer asymmetry points to a data-statistics explanation: the corpus has less distinct information about future years, so the embedding geometry already collapses them; an intervention that adds structured future events to training data should reduce the asymmetry.
  • Read experientially, the result suggests alignment efforts should monitor the internal temporal frame itself, since a model that treats a different year as 'now' could systematically misjudge the relevance and consequences of past and future events.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies temporal cognition in LLMs using a similarity judgment task over years 1525-2524 in 12 models, with a number-to-number control. It reports that larger models' year-similarity judgments are better predicted by a reference-log-linear distance (dref), which assumes a reference point (fixed to 2025) and logarithmic compression, than by log-linear or Levenshtein distances; this is interpreted as evidence that LLMs spontaneously form a subjective temporal reference point and obey the Weber-Fechner law. The paper then presents three supporting analyses: identification of temporal-preferential neurons whose average activations dip near the reference and fit a log-distance regression; linear probes showing a layer-wise shift from numerical to reference-centered temporal representations; and embedding-model analyses suggesting the training corpus itself has a non-linear temporal structure. It concludes with an experientialist perspective and implications for AI alignment.

Significance. If the behavioral result is robust, the paper is a valuable contribution to the emerging cognitive-science analysis of LLMs. Its strengths include a large and systematic dataset (twelve models, one million pairwise judgments), a clean control condition (number vs year), multiple converging levels of analysis, and publicly available code. The multi-level design, from behavior to neurons to corpus structure, is ambitious and goes beyond simple behavioral reports. However, the central inference depends on a fixed, experimenter-chosen reference point and on small R2 differences without uncertainty quantification, so the headline claim needs substantially stronger statistical support before it can be accepted.

major comments (3)
  1. [Methods, 'Reference-Log-Linear distance'; Table 3; Figure 9] The central behavioral evidence fixes R=2025 for all models, but the stimulus range 1525-2524 has midpoint 2024.5, so R=2025 coincides with the center of the presented distribution; for pairs on the same side of R, dref is exactly a log-linear distance from the range center. The paper's own non-parametric sliding-window estimates (Figure 9) place the reference near 2010-2024 across models, and the text acknowledges these estimates are confounded, yet the regressions in Table 3 use the fixed origin. Because the dref advantage is the basis for the 'spontaneous reference point' conclusion, the manuscript should report per-model R values (e.g., each model's sliding-window estimate or an optimized-origin comparison) and a sensitivity analysis over R; otherwise the fit may reflect the experimenter's choice of origin rather than the model's subjective present.
  2. [Results, Table 3] The R2 differences that drive the conclusion are small and no uncertainty is reported; for example, Qwen2.5-72B is 0.3145 vs 0.2775 for dref vs dlog, Qwen2.5-7B is 0.4183 for dref but 0.4373 for dlev, and Gemini-2.0-flash favors dlog (0.4808) over dref (0.4533). Since the one-million pairwise similarity judgments are highly non-independent, raw R2 gaps are not meaningful without block bootstrap or cross-validated intervals and a complexity-adjusted comparison. This is load-bearing because the scaling claim ('larger models spontaneously establish...') is inferred from these gaps.
  3. [Neural Coding, Figure 4] The layer-wise logarithmic fits regress activations against log(|2025-x|) with the same fixed reference, even though the activation troughs in the same figure are the appropriate model-specific reference points. Reporting R2 for a fixed origin chosen by the experimenter cannot distinguish a model-internal logarithmic code from a forced log-distance-to-2025 fit; the authors should fit log(|R*-x|) with R* from the observed trough and compare fits across candidate references. The selection thresholds (Cohen's d > 2.0, FDR p < 0.0001, consistency > 0.95) are also arbitrary and should be varied.
minor comments (5)
  1. [Methods, 'Reference-Log-Linear distance'] The notation dref(i,j) = |log(|R-i|) ◦ log(|R-j|)| is ambiguous; the two cases (same side vs opposite sides of R) should be written out explicitly as separate equations.
  2. [Abstract and Results] The word 'spontaneously' overstates what the evidence shows, since the reference point used in the main regressions was imposed by the authors rather than estimated from each model; the discussion should be reworded to reflect that the metric assumes a reference.
  3. [Table 2] For the number-to-number control, applying dref with R=2025 is a strange baseline because 2025 has no natural role for pure numbers; this makes the comparison in Table 2 harder to interpret.
  4. [Information Exposure] Calling the embedding models 'independent' is only partially justified; they are also text-trained models, so the analysis shows a correlation between two neural text models rather than direct evidence about the underlying corpus. The causal language ('provides the raw material') should be tempered.
  5. [Figure 9] The sliding-window reference-point estimates use window size 5 without robustness checks or confidence intervals; the sensitivity of the estimates to this choice should be reported.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the reference-log-linear distance is tested against observed judgments with a fixed, a priori origin, not fitted to them, and independent sliding-window and neuron-trough analyses support the reference-point claim.

full rationale

The central derivation is not circular. The Reference-Log-Linear distance dref(i,j)=|log(|R-i|) ∘ log(|R-j|)| with R fixed at 2025 is a predictor constructed from the Weber-Fechner hypothesis, but the regression target is the model's observed similarity judgment dLLM = 1 - sLLM, which is not defined in terms of dref or R. The better R2 for dref in Table 3 is therefore an empirical comparison among three theoretical distance metrics, not a tautology. The paper explicitly declines to fit R to the judgments ('treating R as a free parameter for statistical optimization would be insufficient... we fixed 2025 as the reference point'), so the reference point is not a fitted parameter relabeled as a discovery. Independent support comes from the non-parametric diagonal sliding-window estimates (Figure 9) and the neuron-activation trough, neither of which uses dref. The manuscript itself flags that the sliding-window year estimates are 'influenced by other confounding factors' and vary across models (2010-2024) while R is fixed at 2025; this is a validity limitation for the strength of the 'spontaneous reference point' inference, but it does not make any equation equivalent to its input by construction. The self-citations (Li et al. 2025 in the introduction; Li and Li 2025 in the discussion) are peripheral framing and are not load-bearing for the empirical derivation. Overall circularity is minimal.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The central claim rests on three domain assumptions: that year and number prompt differences isolate temporal processing; that embeddings from three embedding models stand in for the training corpus; and that 1-s similarities are linear distances. The main hand-chosen parameters are the fixed reference R=2025 and the neuron-selection thresholds. No new physical entities are introduced; the temporal-preferential neurons and subjective reference point are internal constructs validated only within the paper's own measurements.

free parameters (3)
  • Reference point R in dref = 2025 (fixed by authors)
    The behavioral test of Weber-Fechner temporal compression uses dref, which is defined around R=2025. The paper fixes R rather than fitting it, and model-specific sliding-window estimates range from 2010 to 2024. The choice of 2025 directly shapes the R2 comparisons in Table 3.
  • Neuron selection thresholds = Cohen's d > 2.0; FDR p < 0.0001; consistency > 0.95; top 1000 neurons
    These thresholds define temporal-preferential neurons. The existence and properties of the neural mechanism are conditional on this particular filter, which is not justified by external criteria.
  • Sliding window size = 5
    The non-parametric reference point estimate uses a window size of 5; different window sizes would shift the estimated reference years.
assumptions (3)
  • domain assumption Activation differences between 'Year: x-x-x-x' and 'Number: x-x-x-x' isolate temporal processing.
    Neurons classified as temporal-preferential are those with larger activations for the year framing than the number framing; confounding semantic differences between the words are assumed away. Invoked in Neural Coding.
  • domain assumption Pre-trained embedding models reflect the statistical structure of the LLMs' training corpus.
    The information-exposure analysis measures the semantic similarity of year strings with three embedding models, then treats that similarity as a proxy for the training corpus. Embeddings are themselves model-derived and may share training data and biases with the tested LLMs. Invoked in Information Exposure.
  • domain assumption Similarity ratings can be converted monotonically to psychological distance d=1-s and modeled by linear regression.
    The entire quantitative analysis assumes dLLM = 1-sLLM is a valid distance and that a linear relation to theoretical distances is the right comparison. This is standard in psychophysics but is not verified for LLM outputs.
invented entities (2)
  • Subjective temporal reference point
    purpose: Postulated internal 'present' around which temporal similarity is organized; central to the Weber-Fechner claim.
    Quantified only through the fixed R in dref and through sliding-window minima; no independent measurement outside the paper's own tasks validates it as an internal state.
  • Temporal-preferential neurons
    purpose: A subpopulation of FFN neurons claimed to specifically encode temporal information and logarithmic distance to the reference.
    Identified using paper-specific thresholds and not validated against external benchmarks; the set of neurons is defined by the same data used to characterize it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Other Mind: How Language Models Exhibit Human Temporal Cognition." pith.science (2026). https://pith.science/paper/O36PK7SZ

@misc{pith2026250715851,
  author       = {Pith},
  title        = {Pith review of: The Other Mind: How Language Models Exhibit Human Temporal Cognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O36PK7SZ}},
  note         = {Machine review of arXiv:2507.15851}
}
read the original abstract

As Large Language Models (LLMs) continue to advance, they exhibit certain cognitive patterns similar to those of humans that are not directly specified in training data. This study investigates this phenomenon by focusing on temporal cognition in LLMs. Leveraging the similarity judgment task, we find that larger models spontaneously establish a subjective temporal reference point and adhere to the Weber-Fechner law, whereby the perceived distance logarithmically compresses as years recede from this reference point. To uncover the mechanisms behind this behavior, we conducted multiple analyses across neuronal, representational, and informational levels. We first identify a set of temporal-preferential neurons and find that this group exhibits minimal activation at the subjective reference point and implements a logarithmic coding scheme convergently found in biological systems. Probing representations of years reveals a hierarchical construction process, where years evolve from basic numerical values in shallow layers to abstract temporal orientation in deep layers. Finally, using pre-trained embedding models, we found that the training corpus itself possesses an inherent, non-linear temporal structure, which provides the raw material for the model's internal construction. In discussion, we propose an experientialist perspective for understanding these findings, where the LLMs' cognition is viewed as a subjective construction of the external world by its internal representational system. This nuanced perspective implies the potential emergence of alien cognitive frameworks that humans cannot intuitively predict, pointing toward a direction for AI alignment that focuses on guiding internal constructions. Our code is available at https://TheOtherMind.github.io.

Figures

Figures reproduced from arXiv: 2507.15851 by the authors.

Figure 1
Figure 1. An experientialist perspective of LLMs human [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pair-wise similarities from year 1525 to year 2524 across 12 models with varying sizes [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Distribution of temporal-preferential neurons [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Upper: mean activations of top 1000 temporal preferential neurons to one thousand years from 1525 to 2524 and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Layer-wise performance (R2 ) of linear probes for Log-Linear distance (circle), Reference-Log-Linear distance (square), and Levenshtein distance (triangle) across 10 models [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Pair-wise cosine similarity matrices between em [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Prompts for similarity judgment tasks for year (left) and number (right) [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Pair-wise similarities from Year 1525 to Year 2524 (upper) and Number 1525 to Number 2524 (lower) across 12 [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Non-parametric estimation of subjective temporal reference points using a sliding diagonal window analysis [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Temporal Preference Concepts and their Functions in a Large Language Model

    cs.LG 2026-05 unverdicted novelty 6.5 of 10

    Temporal preference in Qwen3-4B-Instruct-2507 localizes to layers 17–35 (especially L24 attention), has curved residual-stream geometry, is behaviorally unstable, and can be bidirectionally steered.

Reference graph

Works this paper leans on

74 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Alain, G.; and Bengio, Y. 2016. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644

    arXiv 2016

  2. [2]

    Bai, Y.; Kadavath, S.; Kundu, S.; Askell, A.; Kernion, J.; Jones, A.; Chen, A.; Goldie, A.; Mirhoseini, A.; McKinnon, C.; et al. 2022. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073

    arXiv 2022

  3. [3]

    Bengio, Y.; Cohen, M.; Fornasiere, D.; Ghosn, J.; Greiner, P.; MacDermott, M.; Mindermann, S.; Oberman, A.; Richardson, J.; Richardson, O.; et al. 2025 a . Superintelligent agents pose catastrophic risks: Can scientist ai offer a safer path? arXiv preprint arXiv:2502.15657

    arXiv 2025

  4. [4]

    Bengio, Y.; Mindermann, S.; Privitera, D.; Besiroglu, T.; Bommasani, R.; Casper, S.; Choi, Y.; Fox, P.; Garfinkel, B.; Goldfarb, D.; et al. 2025 b . International AI Safety Report. arXiv preprint arXiv:2501.17805

    arXiv 2025

  5. [5]

    Benjamini, Y.; and Hochberg, Y. 1995. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57(1): 289--300

    work page 1995

  6. [6]

    Berti, L.; Giorgi, F.; and Kasneci, G. 2025. Emergent Abilities in Large Language Models: A Survey. arXiv preprint arXiv:2503.05788

    arXiv 2025

  7. [7]

    Binz, M.; and Schulz, E. 2023. Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences, 120(6): e2218523120

    2023

  8. [8]

    Bullmore, E.; and Sporns, O. 2009. Complex brain networks: graph theoretical analysis of structural and functional systems. Nature reviews neuroscience, 10(3): 186--198

    2009

Show all 74 references
  1. [9]

    Chalmers, D. J. 2023. Could a large language model be conscious? arXiv preprint arXiv:2303.07103

    2023 arXiv

  2. [10]

    Chen, J.; Wei, Z.; Ren, Z.; Li, Z.; and Zhang, J. 2025. LR ^ 2 Bench: Evaluating Long-chain Reflective Reasoning Capabilities of Large Language Models via Constraint Satisfaction Problems. arXiv preprint arXiv:2502.17848

    2025 arXiv

  3. [11]

    Clark, A. 1998. Being there: Putting brain, body, and world together again. MIT press

    1998

  4. [12]

    Dalrymple, D.; Skalse, J.; Bengio, Y.; Russell, S.; Tegmark, M.; Seshia, S.; Omohundro, S.; Szegedy, C.; Goldhaber, B.; Ammann, N.; et al. 2024. Towards guaranteed safe ai: A framework for ensuring robust and reliable ai systems. arXiv preprint arXiv:2405.06624

    2024 arXiv

  5. [13]

    L.; and Sireci, S

    Davison, M. L.; and Sireci, S. G. 2000. Multidimensional scaling. In Handbook of applied multivariate statistics and mathematical modeling, 323--352. Elsevier

    2000

  6. [14]

    Dehaene, S. 2003. The neural basis of the Weber--Fechner law: a logarithmic mental number line. Trends in cognitive sciences, 7(4): 145--147

    2003

  7. [15]

    K.; Binz, M.; and Schulz, E

    Demircan, C.; Saanum, T.; Jagadish, A. K.; Binz, M.; and Schulz, E. 2024. Sparse autoencoders reveal temporal difference learning in large language models. arXiv preprint arXiv:2410.01280

    2024 arXiv

  8. [16]

    Dennett, D. C. 1993. Consciousness explained. Penguin uk

    1993

  9. [17]

    A.; Selen, L

    Faisal, A. A.; Selen, L. P.; and Wolpert, D. M. 2008. Noise in the nervous system. Nature reviews neuroscience, 9(4): 292--303

    2008

  10. [18]

    Fechner, G. T. 1948. Elements of psychophysics, 1860

    1948

  11. [19]

    Field, D. J. 1994. What is the goal of sensory coding? Neural computation, 6(4): 559--601

    1994

  12. [20]

    Ganguli, D.; Lovitt, L.; Kernion, J.; Askell, A.; Bai, Y.; Kadavath, S.; Mann, B.; Perez, E.; Schiefer, N.; Ndousse, K.; et al. 2022. Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned. arXiv preprint arXiv:2209.07858

    2022 arXiv

  13. [21]

    K.; et al

    Goldstein, A.; Ham, E.; Schain, M.; Nastase, S.; Zada, Z.; Dabush, A.; Aubrey, B.; Gazula, H.; Feder, A.; Doyle, W. K.; et al. 2023. The temporal structure of language processing in the human brain corresponds to the layered hierarchy of deep language models. arXiv preprint ar...

    2023 arXiv

  14. [22]

    A.; Feder, A.; Emanuel, D.; Cohen, A.; et al

    Goldstein, A.; Zada, Z.; Buchnik, E.; Schain, M.; Price, A.; Aubrey, B.; Nastase, S. A.; Feder, A.; Emanuel, D.; Cohen, A.; et al. 2020. Thinking ahead: spontaneous prediction in context as a keystone of language in humans and machines. BioRxiv, 2020--12

    2020

  15. [23]

    Greenblatt, R.; Denison, C.; Wright, B.; Roger, F.; MacDiarmid, M.; Marks, S.; Treutlein, J.; Belonax, T.; Chen, J.; Duvenaud, D.; et al. 2024. Alignment faking in large language models. arXiv preprint arXiv:2412.14093

    2024 arXiv

  16. [24]

    Guo, H.; Zhang, L.; Feng, X.; and Zheng, Q. 2024. A Review of the Application of Prompt Engineering in the Safety of Large Language Models. In Proceedings of the 2024 2nd International Conference on Information Education and Artificial Intelligence, 424--430

    2024

  17. [25]

    Hahn, M.; and Goyal, N. 2023. A theory of emergent in-context learning as implicit structure induction. arXiv preprint arXiv:2303.07971

    2023 arXiv

  18. [26]

    Han, Y.; Xu, L.; Chen, S.; Zou, D.; and Lu, C. 2024. Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension Ability. arXiv preprint arXiv:2411.19456

    2024 arXiv

  19. [27]

    He, Q.; Zeng, J.; Huang, W.; Chen, L.; Xiao, J.; He, Q.; Zhou, X.; Liang, J.; and Xiao, Y. 2024. Can large language models understand real-world complex instructions? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 18188--18196

    2024

  20. [28]

    Hinton, G. 2024. Will digital intelligence replace biological intelligence. Romanes Lecture, Oxford, UK, 19

    2024

  21. [29]

    R.; Ewart, A.; and Sharkey, L

    Huben, R.; Cunningham, H.; Smith, L. R.; Ewart, A.; and Sharkey, L. 2023. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations

    2023

  22. [30]

    Itzhak, I.; Stanovsky, G.; Rosenfeld, N.; and Belinkov, Y. 2024. Instructed to bias: Instruction-tuned language models exhibit emergent cognitive bias. Transactions of the Association for Computational Linguistics, 12: 771--785

    2024

  23. [31]

    R.; and Bergen, B

    Jones, C. R.; and Bergen, B. K. 2025. Large language models pass the turing test. arXiv preprint arXiv:2503.23674

    2025 arXiv

  24. [32]

    T.; Nam, A.; Sucholutsky, I.; Veselovsky, V.; et al

    Ku, A.; Campbell, D.; Bai, X.; Geng, J.; Liu, R.; Marjieh, R.; McCoy, R. T.; Nam, A.; Sucholutsky, I.; Veselovsky, V.; et al. 2025. Using the tools of cognitive science to understand large language models at different levels of analysis. arXiv preprint arXiv:2503.13401

    2025

  25. [33]

    Kuo, M.; Zhang, J.; Ding, A.; Wang, Q.; DiValentin, L.; Bao, Y.; Wei, W.; Li, H.; and Chen, Y. 2025. H-cot: Hijacking the chain-of-thought safety reasoning mechanism to jailbreak large reasoning models, including openai o1/o3, deepseek-r1, and gemini 2.0 flash thinking. arXiv ...

    2025 arXiv

  26. [34]

    Lakoff, G.; and Johnson, M. 2008. Metaphors we live by. University of Chicago press

    2008

  27. [35]

    Laughlin, S. 1981. A simple coding procedure enhances a neuron's information capacity. Zeitschrift f \"u r Naturforschung c , 36(9-10): 910--912

    1981

  28. [36]

    H.; Li, Z.; Chen, K.; Vera, H

    Lee, J.; Chen, F.; Dua, S.; Cer, D.; Shanbhogue, M.; Naim, I.; \'A brego, G. H.; Li, Z.; Chen, K.; Vera, H. S.; et al. 2025. Gemini embedding: Generalizable embeddings from gemini. arXiv preprint arXiv:2503.07891

    2025 arXiv

  29. [37]

    I.; et al

    Levenshtein, V. I.; et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, volume 10, 707--710. Soviet Union

    1966

  30. [38]

    Li, L.; and Li, C. 2025. Formalizing Lacanian psychoanalysis through the free energy principle. Frontiers in Psychology, 16: 1574650

    2025

  31. [39]

    Li, L.; Wang, Y.; Zhao, H.; Kong, S.; Teng, Y.; Li, C.; and Wang, Y. 2025. Reflection-Bench: Evaluating Epistemic Agency in Large Language Models. In Forty-second International Conference on Machine Learning

    2025

  32. [40]

    Lindsey, J.; Gurnee, W.; Ameisen, E.; Chen, B.; Pearce, A.; Turner, N. L.; Citro, C.; Abrahams, D.; Carter, S.; Hosmer, B.; Marcus, J.; Sklar, M.; Templeton, A.; Bricken, T.; McDougall, C.; Cunningham, H.; Henighan, T.; Jermyn, A.; Jones, A.; Persic, A.; Qi, Z.; Thompson, T. B...

    2025

  33. [41]

    J.; Sucholutsky, I.; Lombrozo, T.; and Griffiths, T

    Liu, R.; Geng, J.; Wu, A. J.; Sucholutsky, I.; Lombrozo, T.; and Griffiths, T. L. 2024. Mind your step (by step): Chain-of-thought can reduce performance on tasks where thinking makes humans worse. arXiv preprint arXiv:2410.21333

    2024 arXiv

  34. [42]

    J.; and Trope, Y

    Maglio, S. J.; and Trope, Y. 2019. Temporal orientation. Current opinion in psychology, 26: 62--66

    2019

  35. [43]

    L.; and Sucholutsky, I

    Marjieh, R.; Veselovsky, V.; Griffiths, T. L.; and Sucholutsky, I. 2025. What is a Number, That a Large Language Model May Know It? arXiv preprint arXiv:2502.01540

    2025 arXiv

  36. [44]

    S.; and Pitts, W

    McCulloch, W. S.; and Pitts, W. 1943. A logical calculus of the ideas immanent in nervous activity. The bulletin of mathematical biophysics, 5: 115--133

    1943

  37. [45]

    A.; Bickel, S.; Mehta, A

    Mischler, G.; Li, Y. A.; Bickel, S.; Mehta, A. D.; and Mesgarani, N. 2024. Contextual feature extraction hierarchies converge in large language models and the brain. Nature Machine Intelligence, 1--11

    2024

  38. [46]

    Muennighoff, N.; Tazi, N.; Magne, L.; and Reimers, N. 2022. MTEB: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316

    2022 arXiv

  39. [47]

    OpenAI. 2024. text-embedding-3-large, https://platform.openai.com/docs/models/text-embedding-3-large. Technical report

    2024

  40. [48]

    Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 27730--27744

    2022

  41. [49]

    Parr, T.; Pezzulo, G.; and Friston, K. J. 2022. Active inference: the free energy principle in mind, brain, and behavior. MIT Press

    2022

  42. [50]

    Pearl, J.; and Mackenzie, D. 2018. The book of why: the new science of cause and effect. Basic books

    2018

  43. [51]

    T.; Muller, D

    Piantadosi, S. T.; Muller, D. C.; Rule, J. S.; Kaushik, K.; Gorenstein, M.; Leib, E. R.; and Sanford, E. 2024. Why concepts are (probably) vectors. Trends in Cognitive Sciences, 28(9): 844--856

    2024

  44. [52]

    Qin, Y.; Hu, S.; Lin, Y.; Chen, W.; Ding, N.; Cui, G.; Zeng, Z.; Zhou, X.; Huang, Y.; Xiao, C.; et al. 2024. Tool learning with foundation models. ACM Computing Surveys, 57(4): 1--40

    2024

  45. [53]

    QwenTeam. 2025. Qwen3-Embedding

    2025

  46. [54]

    Rosenblatt, F. 1958. The perceptron: a probabilistic model for information storage and organization in the brain. Psychological review, 65(6): 386

    1958

  47. [55]

    Seth, A. K. 2024. Conscious artificial intelligence and biological naturalism. Behavioral and Brain Sciences, 1--42

    2024

  48. [56]

    Shani, C.; Jurafsky, D.; LeCun, Y.; and Shwartz-Ziv, R. 2025. From tokens to thoughts: How LLMs and humans trade compression for meaning. arXiv preprint arXiv:2505.17117

    2025

  49. [57]

    Shepard, R. N. 1980. Multidimensional scaling, tree-fitting, and clustering. Science, 210(4468): 390--398

    1980

  50. [58]

    Shojaee, P.; Mirzadeh, I.; Alizadeh, K.; Horton, M.; Bengio, S.; and Farajtabar, M. 2025. The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity. arXiv preprint arXiv:2506.06941

    2025 arXiv

  51. [59]

    W.; Albergo, D.; Borghini, G.; Pansardi, O.; Scaliti, E.; Gupta, S.; Saxena, K.; Rufo, A.; Panzeri, S.; Manzi, G.; et al

    Strachan, J. W.; Albergo, D.; Borghini, G.; Pansardi, O.; Scaliti, E.; Gupta, S.; Saxena, K.; Rufo, A.; Panzeri, S.; Manzi, G.; et al. 2024. Testing theory of mind in large language models and humans. Nature Human Behaviour, 8(7): 1285--1295

    2024

  52. [60]

    Su, J.; Lang, Y.; and Chen, K.-Y. 2023. Can ai solve newsvendor problem without making biased decisions? a behavioral experimental study. A Behavioral Experimental Study (September 1, 2023)

    2023

  53. [61]

    R.; Ziaee, A.; and Nguyen, M

    Suri, G.; Slater, L. R.; Ziaee, A.; and Nguyen, M. 2024. Do large language models show decision heuristics similar to humans? A case study using GPT-3.5. Journal of Experimental Psychology: General, 153(4): 1066

    2024

  54. [62]

    Tang, Z.; and Kejriwal, M. 2024. Humanlike Cognitive Patterns as Emergent Phenomena in Large Language Models. arXiv preprint arXiv:2412.15501

    2024 arXiv

  55. [63]

    B.; and Griffiths, T

    Tenenbaum, J. B.; and Griffiths, T. L. 2001. Generalization, similarity, and Bayesian inference. Behavioral and brain sciences, 24(4): 629--640

    2001

  56. [64]

    Valmeekam, K.; Marquez, M.; Olmo, A.; Sreedharan, S.; and Kambhampati, S. 2023. Planbench: An extensible benchmark for evaluating large language models on planning and reasoning about change. Advances in Neural Information Processing Systems, 36: 38975--38987

    2023

  57. [65]

    Wei, J.; Tay, Y.; Bommasani, R.; Raffel, C.; Zoph, B.; Borgeaud, S.; Yogatama, D.; Bosma, M.; Zhou, D.; Metzler, D.; et al. 2022 a . Emergent abilities of large language models. arXiv preprint arXiv:2206.07682

    2022 arXiv

  58. [66]

    V.; Zhou, D.; et al

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022 b . Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824--24837

    2022

  59. [67]

    I.; Onizuka, M.; Tang, S.; and Xiao, C

    Wu, Z.; Peng, R.; Zheng, S.; Liu, Q.; Han, X.; Kwon, B. I.; Onizuka, M.; Tang, S.; and Xiao, C. 2024. Shall we team up: Exploring spontaneous cooperation of competing llm agents. arXiv preprint arXiv:2402.12327

    2024 arXiv

  60. [68]

    Xu, N.; Zhang, Q.; Du, C.; Luo, Q.; Qiu, X.; Huang, X.; and Zhang, M. 2025. Human-like conceptual representations emerge from language prediction. arXiv preprint arXiv:2501.12547

    2025

  61. [69]

    Yang, Z.; Dong, L.; Du, X.; Cheng, H.; Cambria, E.; Liu, X.; Gao, J.; and Wei, F. 2022. Language models as inductive reasoners. arXiv preprint arXiv:2212.10923

    2022 arXiv

  62. [70]

    Zhong, J.; Shen, W.; Li, Y.; Gao, S.; Lu, H.; Chen, Y.; Zhang, Y.; Zhou, W.; Gu, J.; and Zou, L. 2025. A comprehensive survey of reward models: Taxonomy, applications, challenges, and future. arXiv preprint arXiv:2504.12328

    2025 arXiv

  63. [71]

    Zhu, J.-Q.; and Griffiths, T. L. 2024. Eliciting the priors of large language models using iterated in-context learning. arXiv preprint arXiv:2406.01860

    2024 arXiv

  64. [72]

    Zou, A.; Phan, L.; Chen, S.; Campbell, J.; Guo, P.; Ren, R.; Pan, A.; Yin, X.; Mazeika, M.; Dombrowski, A.-K.; et al. 2023. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405

    2023 arXiv

  65. [73]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  66. [74]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.