Pith. sign in

REVIEW 4 major objections 6 minor 82 references

How Do Artificial Intelligences Think? The Three Mathematico-Cognitive Factors of Categorical Segmentation Operated by Synthetic Neurons

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read This paper claims that a formal neuron's categorical segmentation of the token universe in a large language model is set by three 'mathematico-cognitive factors' that live inside the neuron's aggregation function.

desk verdict The three 'factors' are just the terms of the weighted-sum aggregation function, and the near-unity correlations that 'confirm' them are largely built into the top-100/top-10 selection procedure; only the qualitative taxonomy adds something worth keeping. read the letter →

arxiv 2501.06196 v1 pith:T2NSFQVL submitted 2024-12-26 q-bio.NC cs.AIcs.NE

classification q-bio.NCcs.AIcs.NE
keywords mechanisticinterpretabilitycategoricalsegmentationsyntheticneuronsaggregationfunctionprimingattentionphasingGPT-2XL
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 claims that a formal neuron's categorical segmentation of the token universe in a large language model is set by three 'mathematico-cognitive factors' that live inside the neuron's aggregation function. The factors are categorical priming (the activation levels arriving from precursor neurons), categorical attention (the connection weights attached to those activations), and categorical phasing (the additive summing that co-activates a token across several precursors). The authors test the claim on the first two layers of GPT-2XL, correlating token activation ranks with each factor; they report Spearman correlations of .94 for priming, .999 for attention, and .976 to .989 for phasing, alongside qualitative examples of how precursor categories contribute sub-dimensions to successor categories. If the claim is right, neuron-level semantics can be read directly off the weighted-sum components, and synthetic categories become analyzable with the same vocabulary used for human attention, priming, and conceptual combination.

What carries the argument

The central object is the aggregation function $\sum_{i,j}(w_{i,j}x_{i,j}) + a$ that every formal neuron applies to its inputs, where $x_{i,j}$ are the activation values arriving from precursor neurons, $w_{i,j}$ are the connection weights, and $a$ is a bias. The argument works by identifying each algebraic component with a cognitive factor: the precursor activations give priming, the weights give attention, and the additive summation gives phasing. Around this identity the paper builds its operational observables: for each of 6,400 destination neurons in layer 1 it takes the 10 precursor neurons with the largest connection weights and the 100 tokens with the highest average activation ('core-tokens'), then measures rank correlations between token activation and each factor, and performs a multiple linear regression of activation rank on the three factor variables.

What would settle it

Recompute the same Spearman correlations using all 6,400 precursor neurons of each layer-1 destination neuron (or a random sample of the non-top-10 precursors) and the full token vocabulary instead of the top 100 core-tokens; if the correlations fall well below the reported values, the three-factor story is an artifact of subset selection. A second check is to ablate the top-10 connection weights to zero and observe whether the activation-rank changes match the exact weights-predicted ordering.

Watch

Extended reading notes

Core claim

The central claim, stated in the paper's overview of its three factors, is that the aggregation function $\sum(w_{ij}x_{ij})+a$ genetically shapes the categorical segmentation performed by synthetic neurons. Effect X (categorical priming) is the propagation of precursor activation values into the destination neuron; effect W (categorical attention) is the gating and extraction performed by connection weights; effect $\Sigma$ (categorical phasing) is the additive combination by which a token already core to several precursors arrives with a strengthened activation at the destination. The paper reports that each factor alone shows an extremely strong ordinal relationship with core-token activation rank in layer 1, and that a multiple regression with the three factors yields standardized coefficients $0.86$, $0.56$, and $0.65$ respectively. It concludes that these factors, together with the activation function, constitute the genetic variables of categorical segmentation, operating through categorical complementation and categorical intersection.

Load-bearing premise

The analysis assumes that the only inputs that matter for a neuron's category are its 10 strongest incoming connections and the 100 tokens that already activate it most, so the near-perfect correlations could be an artifact of looking only inside that preselected window.

Editorial extensions

If this is right

  • If the three factors are the generative variables of categorical segmentation, then the semantic content of a neuron can be predicted from a small window of its inputs: the top precursor weights and precursor activations, without probing deeper network states.
  • Because connection weights act as selective extractors of precursor sub-dimensions, interpretability tools can focus on high-weight connections to explain why a particular token enters a neuron's category.
  • Destination categories are compositional: their core-token extensions are assembled by complementation and intersection from precursor categories, so category structure can be studied sub-dimension by sub-dimension.
  • The same factor decomposition should transfer to other transformer layers and models built on the same weighted-sum aggregation, since the identity is a structural property of the architecture.
  • The qualitative examples imply that apparently 'alien' or non-human categories can arise from ordinary statistical co-activation, not from a separate mechanism.

Reading between the lines

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

  • A direct test of the causal reading would be to ablate the top-10 weights or feed in shuffled precursor activations and check whether destination activation rankings shift exactly as the three factors predict; the paper's correlations alone do not establish causation.
  • Because the destination activation is literally a weighted sum of precursor activations, the near-unity correlations may be partly tautological; an independent test would need to compare against a null model that uses random subsets of the remaining 6,390 precursors.
  • The same three-factor lens could be applied to attention heads, biases, layer norms, and residual connections, which the paper excludes; if those components matter, the factor list may need extension.
  • The qualitative complementation examples (e.g., 'manager' plus 'Wenger', 'Klopp') suggest that neuron categories could be used as a probe of statistical category structure in language; one could test whether human category judgments match these synthetic sub-dimensions.
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

4 major / 6 minor

Summary. The paper proposes that the categorical segmentation performed by formal neurons in GPT-2XL is driven by three 'mathematico-cognitive factors' embedded in the aggregation function Σ(w_ij x_ij)+a: categorical priming (effect X, the activation values of precursor neurons), categorical attention (effect W, the connection weights), and categorical phasing (effect Σ, the additive combination). For each of 6,400 layer-1 destination neurons, the authors consider the 100 most activated tokens ('core-tokens') and the 10 highest-weight layer-0 precursor neurons, and they report Spearman correlations between destination-neuron activation ranks and precursor activation or weight variables, along with a multiple regression. They interpret the near-unity correlations as evidence that these three factors drive categorical segmentation. The study is explicitly exploratory and limited to the first two layers of GPT-2XL.

Significance. If the central claim were established, it would offer a bridge between mechanistic interpretability and cognitive-psychological constructs, and the qualitative examples of 'categorical complementation' and 'categorical phasing' are suggestive for hypothesis generation. However, the empirical support is largely tautological: the destination neuron's activation is, by definition, the weighted sum of precursor activations, so correlations between that activation and functions of those same precursor activations and weights are expected to be high even under a null model. The paper is transparent about the illustrative status of its multiple regression, but it does not supply a null baseline, does not test the independence of the three factors, and conditions on a heavily filtered subset. The result, as presented, does not establish three distinct cognitive factors, and the manuscript needs major revision or re-framing before it can support its title claim.

major comments (4)
  1. [§5.1–5.3, Tables 1, 2, 4, 5] The correlations are largely tautological because the dependent variable is constructed from the same quantities as the predictors. For each token t, the destination-neuron activation is defined as y_j(t) = Σ_i w_ij x_ij(t) + a_j (Section 3). The 'effect X' variable is a sum of a subset of the x_ij(t), the 'effect W' variable is a function of the corresponding w_ij, and the 'effect Σ' variable is a count of non-zero contributions in that same sum. A Spearman correlation between the rank of y_j(t) and an unweighted, truncated version of its own defining terms is not an independent test of a cognitive mechanism; near-unity ρ values would likely persist under random or independent terms. The manuscript acknowledges in §5.1 that activation is 'directly a function' of precursor activations, but it does not provide any null model or baseline (e.g., random weights, shuffled tokens) to show that the observed correlations exceed what the algebraic form alone would produce. These correlations therefore cannot, by themselves, support the existence of three separable factors.
  2. [§4.3 and §5.1] The analysis is computed on a heavily filtered subset, and the filtering conditions on the predictor. For each destination neuron, only the 100 highest-activation tokens (core-tokens) are retained, only the 10 highest-weight precursor neurons are considered, and in the priming and attention analyses only core-tokens 'activated in at least one precursor neuron' are included. Because the core-tokens are selected by high destination activation and then restricted to tokens that are also active in the selected precursors, positive association can be induced even if the variables are independent. The paper does not report how many tokens are excluded by this condition, nor does it rerun the analyses on the full set of tokens or on a random subset of precursor neurons. Without such controls, the high correlations in Tables 1, 2, 4, and 5 may be selection artifacts.
  3. [§5.4, Table 7] The only multivariate analysis, the multiple regression on activation rank as a function of the three factors, is explicitly labeled 'illustrative only' because the normality and collinearity assumptions are violated. The standardized coefficients (s-a1 = .86, s-a2 = .56, s-a3 = .65) are therefore not reliable estimates of the relative contributions of the three factors. Since the three variables are constructed from the same aggregation function and are 'likely highly correlated' (as the authors note), the regression cannot establish that priming, attention, and phasing are independent or comparable drivers of categorical segmentation. The paper's central claim in §5.4 that the three factors 'drive the categorical segmentation that neurons perform' is not supported by the quantitative evidence presented.
  4. [§5.1–5.3, definitions of X, W, Σ] The three proposed factors are not separable empirical causes; they are components of an algebraic identity. Effect X is the set of precursor activations x_ij, effect W is the set of weights w_ij, and effect Σ is the summation operation itself. Since y_j = Σ(w_ij x_ij)+a by definition, saying that X, W, and Σ 'drive' y_j is like saying that the factors of a product determine the product. The cognitive interpretation of these algebraic components may be a useful expository frame, but the empirical tests do not distinguish between the claim that the aggregation function determines the activation (a tautology) and the claim that specific cognitive mechanisms with distinct psychological analogues are operative. The paper needs to articulate what would falsify the three-factor model, e.g., by showing that a neuron whose activation is a nonlinear function of precursors, or whose category is not captured by the top-10/top-100 cutoffs, still exhibits the same 'factors.'
minor comments (6)
  1. [§4.2] The phrase 'tokens with the highest average activation values' is ambiguous: the paper does not state over what corpus or context set the average is taken, nor whether activations are averaged over all occurrences of the token. This should be specified for reproducibility.
  2. [§5.4] The text refers to 'Figures 10 to 12' to illustrate residual outliers, but only Figure 10 appears in the manuscript; Figures 11 and 12 are missing.
  3. [Bibliography] The reference list contains multiple entries that are unrelated to the content of the paper (e.g., [33] FEL2024 conference, [34] quantum computing, [36] materials science, [44] psoriasis), and several references are duplicated ([51] and [52] are the same Nadeau entry; [56] and [57] are both self-citations). These should be corrected or removed.
  4. [§3 and §5.3] The third effect is introduced as 'effect P' in Section 3 but later denoted 'effect Σ' in Section 5.3; the notation should be made consistent throughout.
  5. [Figures 2, 4, 7, 10] The example neurons are described as 'control neurons,' but no criterion for selecting these examples is given. State how the control neurons were chosen and whether they are representative or illustrative.
  6. [General] The abstract and title make a broad claim about 'How Do Artificial Intelligences Think?' but the study only examines two layers of GPT-2XL. The conclusions should be explicitly limited to this architecture and layer range.

Circularity Check

5 steps flagged · score 8.0 of 10

The paper's three 'mathematico-cognitive factors' are the components of the aggregation function itself; the near-unity correlations and the central 'drive' claim restate y = Σ(wx)+a by construction.

  1. self definitional [Section 3, Problem Statement (definition of the three effects)]
    "Observing the nature and operators constitutive of this aggregation function, of the form P(wi,jxi,j) +a, suggests that it mathematically generates and formats the categorical segmentation performed by synthetic neurons through at least three mathematico-cognitive factors."

    In Section 3, the three 'factors' are defined as the components of the aggregation function: x (precursor activations), w (weights), and the additive combination Σ. The output activation of the destination neuron is exactly y_j = Σ_i w_ij x_ij + b_j, so saying that these three components 'generate and format' segmentation is a restatement of the definition of y_j. The later empirical sections therefore cannot count as independent evidence for three cognitive causes; there is no alternative mechanism or counterfactual in which those components are absent but y_j is still produced.

  2. self definitional [Section 5.1, Synthetic Categorical Priming (quantitative test; Table 1)]
    "Mathematically, due to the construction of the aggregation function Σ( wi,jxi,j) + a, for a given element (e.g., a token or other), the activation value of the category carried by a destination neuron (on layer n) is directly a function (modulo the activation function) of the activation values xi,j of the categories associated with its precursor neurons (on the subordinate layer n − 1)."

    The priming effect is defined as the x_i,j terms in y = Σ(wx)+a. The test ranks the 100 largest-y tokens and correlates that rank with the sum of x in only the 10 highest-w precursors, using only tokens activated in at least one precursor. Because core-token status is defined by y rank, and x enters y linearly, a high ordinal correlation (ρ=.94) is built into the selection and the defining equation. The phrase 'directly a function' concedes this; the result adds no mechanism beyond the weighted sum.

3 more flagged steps
  1. self definitional [Section 5.2, Synthetic Categorical Attention (Section 5.2.1; Table 2)]
    "This is based on the mathematical construction of the aggregation function Σ( wi,jxi,j) + a, where, for a given element (a token), its activation value within the category associated with a destination neuron is inherently dependent (apart from the activation function) on the connection weights wi,j between this destination neuron and its precursor neurons."

    Attention effect W is the set of weights w in y = Σ(wx)+a. Top-10 precursors are themselves chosen by largest w (Section 4.3), and core-tokens by largest y; correlating y rank with cumulative w over those same precursors is correlating the output with one of its own summands. The observed ρ=.999 is essentially deterministic. The paper's wording 'inherently dependent' on w acknowledges that no independent variation is tested; the correlation is a mathematical identity in disguise.

  2. self definitional [Section 5.3, Synthetic Categorical Phasing (Section 5.3.1; Tables 4-5)]
    "a token, previously highly activated for different precursor neurons (i.e., a core-token of these precursor neurons), must, due to the mathematical construction of the aggregation function, be associated with a high activation level within the related destination neuron."

    The phasing effect Σ is the summation operator itself. The quoted sentence states that a token co-activated in precursors 'must' be highly activated in the destination 'due to the mathematical construction'—that is, the hypothesis is already entailed by y = Σ(wx)+a. Counting co-core-tokens and correlating the count with y rank (ρ≈.98) is an empirical restatement of the same entailment. No data could distinguish this 'effect' from the linear aggregation function that defines the neuron's output.

  3. self definitional [Section 5.4, Overview of the Three Factors (Table 7)]
    "we conducted a multiple linear regression on the activation rank of core-tokens in destination neurons as a function of (i) the average number of times these core-tokens are also core-tokens in the associated precursor neurons (a1) (effect Σ), (ii) the average connection weight of destination neurons with their associated precursor neurons (a2) (effect w), and (iii) the average activation of these core-tokens in the relevant precursor neurons (a3) (effect x)."

    The regression uses activation rank (defined by y) as dependent variable and predictors that are algebraic components or aggregates of the same y: co-core-token count (a proxy for Σ), connection weights w, and precursor activations x. Regressing a quantity on its own parts cannot establish that the parts 'drive' it; it only decomposes the defining formula. The paper itself warns the regression is 'illustrative only' because normality/collinearity assumptions fail, so the standardized coefficients are treated as support for the central claim despite the test being invalid. This is the final step where the three definitional components are presented as confirmed causes.

full rationale

The circularity is not bibliographic: no load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggling citation drives the argument. It is definitional. Section 3 identifies the three factors as the variables and operator of the aggregation function y = Σ(wx)+a: x is priming, w is attention, and the additive sum is phasing. Sections 5.1-5.3 then test each factor by correlating the neuron's activation rank with the same x, w, or co-activation count, computed on subsets (top-100 core-tokens, top-10 weights) that are themselves selected by y and w. The paper's own language—'directly a function', 'inherently dependent', and 'must, due to the mathematical construction'—concedes that these relationships are entailed by the definition of the output. Section 5.4's multiple regression regresses y on its own components and is admitted to be illustrative only because normality and collinearity assumptions fail. No null model or de-correlated baseline is provided, so the near-unity Spearman correlations cannot distinguish the proposed cognitive mechanism from the arithmetic of the aggregation function itself. The central claim that the three factors 'drive' categorical segmentation therefore reduces by construction to the formula that defines the neuron's activation. Score 8 rather than 10 because the paper is explicitly exploratory and transparent about the regression's limitations, but the definitional identification remains load-bearing throughout.

Assumptions & free parameters 3 free parameters · 4 assumptions · 3 invented entities

The central claim rests on the aggregation function being the complete description of a neuron's categorical operation, on the 10 highest-weight precursors and 100 most-activated tokens being the only relevant inputs and outputs, and on the transposition of human cognitive terms. The three effects are not independently evidenced; they are names for terms of the sum and are validated only on the same selected subset. There are no fitted physical constants, but the cutoffs 100 and 10 are hand-chosen parameters that shape every reported correlation.

free parameters (3)
  • core_token_cutoff = 100
    Chosen by hand to define each neuron's category extension; all correlations and the multiple regression are restricted to these tokens (Sections 4.2 and 5.4).
  • precursor_neuron_cutoff = 10
    Chosen by hand; only the 10 highest-weight precursors from layer 0 are included in every analysis (Section 4.3).
  • inclusion_criterion = core-token activated in at least one precursor
    Exclusion rule applied in priming, attention, and phasing analyses; it removes low-activation tokens and inflates positive correlations (Sections 5.1 to 5.3).
assumptions (4)
  • domain assumption The aggregation function Σ(w_ij x_ij)+a, together with the activation function, fully characterizes a formal neuron's categorical operation.
    Invoked throughout Sections 3 to 5; attention heads, biases, layer norms, and residual connections are not modeled.
  • ad hoc to paper The 10 highest-weight precursor neurons are the only inputs that matter for the destination neuron's category.
    Introduced in Section 4.3; without this subsetting, the near-unity correlations are not tested.
  • ad hoc to paper A neuron's category extension is adequately represented by its 100 most activated tokens and their rank.
    Defined in Section 4.2; all hypotheses are tested only on this subset.
  • domain assumption Human cognitive concepts (priming, attention, phasing) can be transposed to the components of artificial neuron aggregation functions.
    Stated in Sections 5.1 to 5.3 as foundational for the naming of the effects.
invented entities (3)
  • Synthetic categorical priming (effect X)
    purpose: Names the influence of precursor neuron activation values on destination neuron activation.
    Only supported by correlation with activation rank, which is definitionally tied to the aggregation function; no external prediction is made.
  • Synthetic categorical attention (effect W)
    purpose: Names the influence of connection weights on category formation and discrimination.
    Correlation with activation rank and shared core-token counts is computed on the same selected subset used to define the effect.
  • Synthetic categorical phasing (effect Σ)
    purpose: Names the additive co-activation of precursor categories as a determiner of destination activation.
    The effect is the summation operation itself; evidence is the rank correlation of output with number of co-activated precursors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Do Artificial Intelligences Think? The Three Mathematico-Cognitive Factors of Categorical Segmentation Operated by Synthetic Neurons." pith.science (2026). https://pith.science/paper/T2NSFQVL

@misc{pith2026250106196,
  author       = {Pith},
  title        = {Pith review of: How Do Artificial Intelligences Think? The Three Mathematico-Cognitive Factors of Categorical Segmentation Operated by Synthetic Neurons},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T2NSFQVL}},
  note         = {Machine review of arXiv:2501.06196}
}
read the original abstract

How do the synthetic neurons in language models create "thought categories" to segment and analyze their informational environment? What are the cognitive characteristics, at the very level of formal neurons, of this artificial categorical thought? Based on the mathematical nature of algebraic operations inherent to neuronal aggregation functions, we attempt to identify mathematico-cognitive factors that genetically shape the categorical reconstruction of the informational world faced by artificial cognition. This study explores these concepts through the notions of priming, attention, and categorical phasing.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

82 extracted references · 23 canonical work pages

  1. [1]

    R., Hansen, M., Iarosz, K

    Protachevicz, P. R., Hansen, M., Iarosz, K. C., Caldas, I. L., Batista, A. M., & Kurths, J. (2021). Emergence of neuronal synchronisation in coupled areas. Frontiers in Computational Neuroscience, 15, 663408. DOI: 10.3389/fncom.2021.663408

  2. [2]

    Schmalzried, M. (2024). The need of a self for self-driving cars: a theoretical model applying homeostasis to self driving. arXiv preprint arXiv:2407.12795. DOI: 10.48550/arXiv.2407.12795

  3. [4]

    G., Lioma, C., & Augenstein, I

    Atanasova, P., Simonsen, J. G., Lioma, C., & Augenstein, I. (2020). Gen- erating Fact Checking Explanations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 7352–7364). Association for Computational Linguistics. DOI: 10.18653/v1/2020.acl- main.656

  4. [5]

    Barkan, R. (2021). The Role of Cognitive Biases in Human Decision Making. Journal of Behavioral Decision Making , 34(3), 243–255. DOI: 10.1002/bdm.2210

  5. [6]

    Barr, W., & Bieliauskas, L. A. (2024). Neuropsychology of Decision Mak- ing: A Clinical Perspective. Neuropsychology Review, 34(1), 1–15. DOI: 10.1007/s11065-023-09500-1

  6. [8]

    Will You Find These Shortcuts ?

    Bastings, J., Ebert, S., Zablotskaia, P., Sandholm, A., & Filippova, K. (2022). “Will You Find These Shortcuts ? ” A Protocol for Evaluating the Faithfulness of Input Salience Methods for Text Classification. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Pro- cessing. https://doi.org/10.18653/v1/2022.emnlp-main.64

  7. [9]

    Bathia, N., & Richie, D. (2024). Advances in Reinforcement Learning: Ap- plications and Challenges. Artificial Intelligence Review , 57(2), 123–145. DOI: 10.1007/s10462-023-10123-4

  8. [10]

    Beaufils, M. (1996). Les r´ eseaux de neurones artificiels: Mod` eles et applications. Revue d’Intelligence Artificielle , 10(4), 365–387. DOI: 10.1016/S0992-499X(97)80001-2

Show all 82 references
  1. [11]

    Bills, S., Cammarata, N., Mossing, D., Saunders, W., Wu, J., Tillman, H., Gao, L., Goh, G., Sutskever, I., & Leike, J. (2023). Language models 26 can explain neurons in language models. OpenAI. https://openaipublic. blob.core.windows.net/neuron-explainer/paper/index.html

  2. [12]

    Bloch, H. (1992). Grand dictionnaire de la psychologie

  3. [13]

    Bolognesi, M. (2020). Where Words Get Their Meaning: Cognitive Pro- cessing and Distributional Modelling of Word Meaning . John Benjamins Publishing Company. DOI: 10.1075/ftl.7

  4. [14]

    Bricken, T., Schaeffer, R., Olshausen, B., & Kreiman, G. (2023). Emergence of Sparse Representations from Noise. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Re- search, 202:3148-3191. Available from https://proceedin...

  5. [16]

    Canales-Johnson, A., Silva, C., Huepe, D., Rivera-Rei, ´A., Noreika, V., Del Carmen Garcia, M., Silva, W., Vaucheret, E., Sede˜ no, L., Couto, B., Melloni, M., Ib´ a˜ nez, A., Chennu, S., Bekinschtein, T. A. (2015). Auditory feedback differentially modulates behavioral and neu...

  6. [19]

    Cowan, N. (2024). Working Memory Capacity: Theories and Applica- tions. Annual Review of Psychology, 75, 1–25. DOI: 10.1146/annurev-psych- 010723-120001

  7. [20]

    Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., & Wei, F. (2022). Knowl- edge Neurons in Pretrained Transformers. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1 : Long Papers). https://doi.org/10.18653/v1/2022.acl-long.581

  8. [21]

    A., & Glass, J

    Dalvi, F., Durrani, N., Sajjad, H., Belinkov, Y., Bau, D. A., & Glass, J. (2019, January). What is one grain of sand in the desert? Analyzing individual neurons in deep NLP models. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence (AAAI, Oral presen...

  9. [22]

    R., Alam, F., Durrani, N., Xu, J., & Saj- jad, H

    Dalvi, F., Khan, A. R., Alam, F., Durrani, N., Xu, J., & Saj- jad, H. (2022). Discovering Latent Concepts Learned in BERT. In International Conference on Learning Representations (ICLR) . DOI: 10.48550/arXiv.2201.10020

  10. [23]

    Danilevsky, M., Qian, K., Aharonov, R., Katsis, Y., Kawas, B., & Sen, P. (2020). A Survey of the State of Explainable AI for Natural Language Pro- cessing. arXiv (Cornell University). https://doi.org/10.48550/arxiv. 2010.00711

  11. [24]

    A., Durrani, N., Sajjad, H., Dalvi, F., & Belinkov, Y

    Dar, S. A., Durrani, N., Sajjad, H., Dalvi, F., & Belinkov, Y. (2023). Prob- ing Pre-trained Language Models for Temporal Knowledge. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguis- tics (ACL). DOI: 10.18653/v1/2023.acl-long.123

  12. [25]

    S., Lee, J

    Du, S. S., Lee, J. D., Li, H., Wang, L., & Zhai, (2019). Gradient descent finds global minima of deep neural networks, 1675-1685

  13. [26]

    F., & Cabi, S

    Du, Y., Konyushkova, K., Denil, M., Raju, A., Landon, J., Hill, F., Nando, D. F., & Cabi, S. (2023). Vision-Language Models as Success Detectors . arXiv (Cornell University). https://doi.org/10.48550/arxiv.2303.07280

  14. [27]

    Duncan, J. (1984). Selective Attention and the Organization of Visual In- formation. Journal of Experimental Psychology: General , 113(4), 501-517. DOI: 10.1037/0096-3445.113.4.501

  15. [28]

    Durrani, N., Sajjad, H., Dalvi, F., & Belinkov, Y. (2022). On the Trans- formation of Latent Space in Fine-Tuned NLP Models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP). DOI: 10.18653/v1/2022.emnlp-main.123

  16. [29]

    Echterhoff, J., Yan, A., Han, K., Abdelraouf, A., Gupta, R., & McAuley, J. (2024). Driving through the Concept Gridlock: Unrav- eling Explainability Bottlenecks in Automated Driving . Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). htt...

  17. [30]

    Enguehard, J. (2023). Extrmask: A Method for Explaining Time Se- ries Predictions by Masking. arXiv preprint arXiv:2301.08552 . DOI: 10.48550/arXiv.2301.08552

  18. [31]

    W., & Keane, M

    Eysenck, M. W., & Keane, M. T. (2020). Cognitive Psychology: A Student’s Handbook (8th ed.). Psychology Press. DOI: 10.4324/9780429449229

  19. [32]

    Fan, Y., Dalvi, F., Durrani, N., & Sajjad, H. (2023). Evaluating Neu- ron Interpretation Methods of NLP Models . arXiv (Cornell University). https://doi.org/10.48550/arxiv.2301.12608 28

  20. [33]

    National Centre for Nuclear Research. (2024). 41st International Free Electron Laser Conference (FEL2024) . Warsaw, Poland. Retrieved from https://fel2024.org/

  21. [34]

    Funayama, T., & Shibata, K. (2024). Advances in Quantum Computing: A Comprehensive Review. Journal of Quantum Information Science , 12(1), 45–67. DOI: 10.4236/jqis.2024.121004

  22. [35]

    Geva, M., Schuster, R., Berant, J., & Levy, O. (2023). Transformer Feed-Forward Layers Are Key-Value Memories. In Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS) . DOI: 10.48550/arXiv.2012.14913

  23. [36]

    Gresch, D., & M¨ uller, K. (2024). Machine Learning in Materials Science: Recent Progress and Emerging Applications. Advanced Materials, 36(5), 2105678. DOI: 10.1002/adma.202105678

  24. [37]

    A., Reicher, S

    Haslam, S. A., Reicher, S. D., & Platow, M. J. (2020). The New Psychology of Leadership: Identity, Influence, and Power (2nd ed.). Routledge. DOI: 10.4324/9781351108225

  25. [38]

    A., & P´ erez-Gonz´ alez, J

    Hern´ andez-Guti´ errez, C. A., & P´ erez-Gonz´ alez, J. (2024). Deep Learning Techniques for Natural Language Processing: A Survey. IEEE Transac- tions on Neural Networks and Learning Systems , 35(2), 1234–1256. DOI: 10.1109/TNNLS.2023.3101234

  26. [39]

    Howell, D. C. (2008). Fundamental Statistics for the Behavioral Sciences (6th ed.). Wadsworth Publishing. DOI: 10.1111/j.1467- 985X.2008.00508 14.x

  27. [40]

    Kandpal, N., Deng, H., Roberts, A., Wallace, E., & Raffel, C. (2023). Large Language Models Struggle to Learn Long-Tail Knowledge . arXiv (Cornell University). https://doi.org/10.48550/arxiv.2211.08411

  28. [41]

    Capuano, F., & Kaup, B. (2024). Pragmatic Reasoning in GPT Mod- els: Replication of a Subtle Negation Effect. Proceedings of the An- nual Meeting of the Cognitive Science Society, 46. Retrieved from https://escholarship.org/uc/item/22q5920s

  29. [42]

    A., Bouadjenek, M

    Kheya, T. A., Bouadjenek, M. R., & Aryal, S. (2024). The Pursuit of Fair- ness in Artificial Intelligence Models: A Survey. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2403.17333

  30. [43]

    Luo, J., Zhuo, W., Liu, S., & Xu, B. (2024). The Opti- mization of Carbon Emission Prediction in Low Carbon En- ergy Economy under Big Data . IEEE Access, 12, 14690-14702. https://doi.org/10.1109/access.2024.3351468 29

  31. [44]

    C., Tsoi, L

    Ma, F., Plazyo, O., Billi, A. C., Tsoi, L. C., Xing, X., Wasikowski, R., Gharaee-Kermani, M., Hile, G., Jiang, Y., Harms, P. W., Xing, E., Kirma, J., Xi, J., Hsu, J., Sarkar, M. K., Chung, Y., Di Dom- izio, J., Gilliet, M., Ward, N. L., et al. (2023). Single cell and spatial s...

  32. [45]

    Marconato, E., & al. (2024). BEARS Make Neuro-Symbolic Models Aware of their Reasoning Shortcuts. arXiv preprint arXiv:2402.12240. DOI: 10.48550/arXiv.2402.12240

  33. [46]

    Marty, P., Romoli, J., Sudo, Y., & Breheny, R. (2024). Implicature priming, salience, and context adaptation. Cognition, 244, 105667. DOI: 10.1016/j.cognition.2023.105667

  34. [48]

    Mitchell, M. (2021). Abstraction and analogy-making in artificial intelli- gence. Annals of the New York Academy of Sciences , 1505(1), 79-101. DOI: 10.1111/nyas.14619

  35. [49]

    J., Johnson, M., & Steed- man, M

    McKenna, N., Li, T., Cheng, L., Hosseini, M. J., Johnson, M., & Steed- man, M. (2023). Sources of Hallucination by Large Language Models on Inference Tasks. arXiv (Cornell University). https://doi.org/10.48550/ arxiv.2305.14552

  36. [50]

    Mousi, B., Durrani, N., & Dalvi, F. (2023). Can LLMs facilitate interpre- tation of pre-trained language models? arXiv preprint arXiv:2305.13386 . DOI: 10.48550/arXiv.2305.13386

  37. [51]

    Nadeau, R. (1999). Vocabulaire technique et analytique de l’´ epist´ emologie. Presses universitaires de France

  38. [52]

    Nadeau, R. (1999). Vocabulaire technique et analytique de l’´ epist´ emologie. Presses Universitaires de France

  39. [53]

    Nanda, N., Lee, A., & Wattenberg, M. (2023). Emergent linear represen- tations in world models of self-supervised sequence models. arXiv preprint arXiv:2309.00941. DOI: 10.48550/arXiv.2309.00941

  40. [54]

    Nosofsky, R. M. (1986). Attention, similarity, and the identifica- tion–categorization relationship. Journal of Experimental Psychology: Gen- eral, 115(1), 39

  41. [55]

    Paolo, G., Gonzalez-Billandon, J., & K´ egl, B. (2024). A call for embodied AI. arXiv preprint arXiv:2402.03824 . DOI: 10.48550/arXiv.2402.03824. 30

  42. [56]

    Pichat, M. (2023). Collaboration des intelligences humaine et artificielle: alignement et psychologie de l’IA. Actes du colloque Intelligence artifi- cielle collaborative & impacts manag´ eriaux au sein des organisations du 30/06/2023 coorganis´ e par l’Universit´ e Paris Daup...

  43. [57]

    Pichat, M. (2024a). Psychologie de l’IA et alignement cognitif. Actes du colloque Intelligence artificielle collaborative, management et d´ eveloppement des organisations du 24/05/2024 coorganis´ e par l’Universit´ e Paris Dauphine-PSL et le Cabinet Chrysippe R&D. Avail- able ...

  44. [58]

    Pichat, M. (2024). Psychology of Artificial Intelligence: Epistemological Markers of the Cognitive Analysis of Neural Networks. arXiv (Cornell Uni- versity). https://doi.org/10.48550/arxiv.2407.09563

  45. [59]

    Posner, M. I. (1978). Chronometric Explorations of Mind . Lawrence Erl- baum Associates

  46. [60]

    I., & Snyder, C

    Posner, M. I., & Snyder, C. R. R. (1975). Attention and Cognitive Control. In R. L. Solso (Ed.), Information Processing and Cognition: The Loyola Symposium (pp. 55-85). Lawrence Erlbaum Associates. DOI: 10.4324/9781315784786

  47. [61]

    Raieli, S., Altahhan, A., Jeanray, N., Gerart, S., & Vachenc, S. (2024). Escaping the Forest: Sparse Interpretable Neural Networks for Tabular Data. arXiv preprint arXiv:2410.17758 . DOI: 10.48550/arXiv.2410.17758

  48. [62]

    Ribary, U., & Ward, L. M. (2024). Synchronization and functional con- nectivity dynamics across TC-CC-CT networks: Implications for clini- cal symptoms and consciousness. In Phenomenological Neuropsychiatry: How Patient Experience Bridges the Clinic with Clinical Neuroscience ...

  49. [63]

    Richard, J. C. (1980). The Language Teaching Matrix . Cambridge Univer- sity Press

  50. [64]

    D., & Love, B

    Roads, B. D., & Love, B. C. (2024). Modeling Similarity and Psy- chological Space. Annual Review of Psychology , 75(1), 215–240. DOI: 10.1146/annurev-psych-040323-115131

  51. [65]

    Rzechorzek, A. (2024). Understanding Cognitive Processes: In- sights from Recent Research. Journal of Cognitive Neuroscience . DOI: 10.1162/jocn a 01678. 31

  52. [66]

    Schneider, W., & Shiffrin, R. M. (1977). Controlled and Automatic Human Information Processing: I. Detection, Search, and Attention. Psychological Review, 84(1), 1-66

  53. [67]

    Shavikloo, M., Esmaeili, A., Valizadeh, A., & Madadi Asl, M. (2024). Syn- chronization of delayed coupled neurons with multiple synaptic connections. Cognitive Neurodynamics, 18(2), 631-643. DOI: 10.1007/s11571-023-10013- 9

  54. [68]

    Tipper, S. P. (1985). The Negative Priming Effect: Inhibitory Priming by Ignored Objects. The Quarterly Journal of Experimental Psychology , 37A(4), 571-590. DOI: 10.1080/14640748508400920

  55. [69]

    Treisman, A., & Gelade, G. (1980). A Feature-Integration Theory of Attention. Cognitive Psychology , 12(1), 97-136. DOI: 10.1016/0010- 0285(80)90005-5

  56. [70]

    R., Hassid, M., Heafield, K., Hooker, S., Raffel, C., Martins, P

    Treviso, M., Lee, J., Ji, T., Van Aken, B., Cao, Q., Ciosici, M. R., Hassid, M., Heafield, K., Hooker, S., Raffel, C., Martins, P. H., Martins, A. F. T., Forde, J. Z., Milder, P., Simpson, E., Slonim, N., Dodge, J., Strubell, E., Balasubramanian, N.,. . . Schwartz, R. (2023). ...

  57. [71]

    Varela, F. (1984). The creative circle. In P. Watzlawick (Ed), The invented reality. London: W W Norton & Co Inc

  58. [72]

    Varela, F. J. (1988). Cognitive Science: A Cartography of Current Ideas . MIT Press.Varela1996

  59. [73]

    Varela, F. J. (1996). Invitation aux sciences cognitives. ´Editions du Seuil eBooks. http://inventin.lautre.net/livres/ Varela-Invitation-aux-sciences-cognitives.pdf

  60. [74]

    Vergnaud, G. (2009). Activit´ e, d´ eveloppement, repr´ esentation. In M. Merri (Ed.), Activit´ e humaine et conceptualisation. Questions ` a G´ erard Vergnaud (pp. 149–154). Presses universitaires du Mirail

  61. [75]

    Vergnaud, G. (2016). Relations entre conceptualisations dans l’action et signifiants langagiers et symboliques. In Symposium latino- am´ ericain de didactique de math´ ematique, Bonito, Br´ esil. Disponible sur : https://www.gerard-vergnaud.org/texts/gvergnaud_2016_ signifiant...

  62. [76]

    Voita, E., Sennrich, R., & Titov, I. (2021). Language modeling, lex- ical translation, reordering: The training process of NMT through the lens of classical SMT. arXiv preprint arXiv:2109.01396 . DOI: 10.48550/arXiv.2109.01396. 32

  63. [77]

    Kokhlikyan, N., Miglani, V., Martin, M., Wang, E., Reynolds, J., Mel- nikov, A., Lunova, N., & Reblitz-Richardson, O. (2020). Captum: A uni- fied and generic model interpretability library for PyTorch. arXiv preprint arXiv:2009.07896. DOI: 10.48550/arXiv.2009.07896

  64. [78]

    Watzlawick, P. (1977). How real is real? London: Vintage Books

  65. [79]

    H., & Fisch, R

    Watzlawick, P., Weakland, J. H., & Fisch, R. (1984). Change: Principles of Problem Formation and Problem Resolution . W. W. Norton & Company. DOI: 10.1002/9781119164894

  66. [80]

    pyOptSparse: A Python framework for large-scale con- strained nonlinear optimization of sparse systems

    Wu et al., (2020). pyOptSparse: A Python framework for large-scale con- strained nonlinear optimization of sparse systems . Journal of Open Source Software, 5(54), 2564. DOI: 10.21105/joss.02564

  67. [81]

    Ji, M., & Wu, Z. (2022). Automatic detection and severity analysis of grape black measles disease based on deep learning and fuzzy logic . Computers and Electronics in Agriculture , 193, 106718

  68. [82]

    Wu, W. (2024). We know what attention is! . Trends in Cognitive Sciences, 28(4), 304-318

  69. [83]

    Xu, W., & Futrell, R. (2024). A hierarchical Bayesian model for syntactic priming. arXiv preprint arXiv:2405.15964 . DOI: 10.48550/arXiv.2405.15964

  70. [84]

    Zadeh, L. A. (1996). Fuzzy Logic = Computing with Words. IEEE Trans- actions on Fuzzy Systems , 4(2), 103-111. DOI: 10.1109/91.493904

  71. [85]

    A., Kirko- rian, H., & Lupyan, G

    Zettersten, M., Bredemann, C., Kaul, M., Ellis, K., Vlach, H. A., Kirko- rian, H., & Lupyan, G. (2024). Nameability supports rule-based category learning in children and adults. Child Development , 95(2), 497-514. DOI: 10.1111/cdev.14008

  72. [86]

    Zheng, Y., & Stewart, N. (2024). Improving EFL students’ cultural aware- ness: Reframing moral dilemmatic stories with ChatGPT. Computers And Education Artificial Intelligence , 6, 100223. https://doi.org/10.1016/ j.caeai.2024.100223

  73. [87]

    Zhao, H., Chen, H., Yang, F., Liu, N., Deng, H., Cai, H., Wang, S., Yin, D., & Du, M. (2023). Explainability for Large Language Models: A Survey. arXiv (Cornell University) . DOI: 10.48550/arxiv.2309.01029

  74. [88]

    & Sun, M

    Zhang, Z., Song, Y., Yu, G., Han, X., Lin, Y., Xiao, C., . . . & Sun, M. (2024). ReLU 2 Wins: Discovering Efficient Activation Functions for Sparse LLMs. arXiv preprint arXiv:2402.03804. DOI: 10.48550/arXiv.2402.03804. 33

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.