Sequence-to-Sequence Models with Attention Mechanistically Map to the Architecture of Human Memory Search

Nikolaus Salvatore; Qiong Zhang

REVIEW 3 major objections 4 minor 1 cited by

Sequence-to-Sequence Models with Attention Mechanistically Map to the Architecture of Human Memory Search

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper maps the attention mechanism of sequence-to-sequence translation models onto the context-reinstatement mechanism of the CMR memory model, making the translation architecture a cognitive model of free recall.

desk verdict The attention-to-CMR mapping is a good idea that needs one technical fix: the equivalence in Eqs. 22–23 does not follow as written. read the letter →

arxiv 2506.17424 v1 pith:PKAULZHK submitted 2025-06-20 q-bio.NC cs.LG

Nikolaus Salvatore , Qiong Zhang This is my paper · ORCID

classification q-bio.NCcs.LG

keywords sequence-to-sequencemodelsattentionmechanismContextMaintenanceandRetrievalfreerecallreinstatementworkingmemoryrationalanalysisneuralmachinetranslation

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 tries to establish that a foundational neural machine translation architecture—an RNN sequence-to-sequence model with dot-product attention—implements the same retrieval machinery as the Context Maintenance and Retrieval (CMR) model of human memory search. The load-bearing identification is that the decoder's hidden state plays the role of CMR's current recall context, the encoder's hidden states play the roles of item encoding contexts, and the attention weights are the same softmax over context similarities that CMR uses to reinstate prior contexts. If this is right, the translation architecture can be used directly as an interpretable cognitive model of free recall: the paper shows it fits individual human serial position curves, first-recall probabilities, and transition probabilities on a 171-subject dataset, and that a reinforcement-learned version converges to the optimal free recall policy previously derived for CMR. The convergence also carries a functional argument: because machine translation architectures were selected for translation performance, their independent discovery of context reinstatement is evidence that the context architecture in CMR serves an adaptive purpose rather than being an arbitrary modeling choice.

What carries the argument

The central object is the identification of the seq2seq encoder as CMR's encoding context and the decoder as CMR's recall context. Concretely, CMR updates context as $c_i = \rho c_{i-1} + \beta x_i$ and stores item-to-context associations $M_{FC}^{\mathrm{exp}}$; at recall, the experimental context of a just-recalled item is $\alpha_j = M_{FC}^{\mathrm{exp}} f_{j-1}$, which in expectation becomes a softmax-weighted sum of encoding contexts. The seq2seq model updates hidden states with gated recurrent units and computes Luong dot-product attention $\alpha_j^{\mathrm{RNN}} = \sum_i w_i^j h_i$ with $w_i^j = \mathrm{softmax}(h_j^\top h_i)$. The paper's move is to equate these two softmax expressions term by term, mapping $h_j$ to $c_j$ and $h_i$ to $c_{i-1}$, so attention weights become context-reinstatement probabilities. This equivalence is what lets the trained network's internal states be read as CMR's context representations.

What would settle it

Train the seq2seq model on the free-recall task, then, trial by trial, compute its attention weights $w_i^j$ and compare them with CMR's context-reinstatement probabilities $\exp(k c_{i-1}^\top c_{j-1})/\sum_{i'} \exp(k c_{i'-1}^\top c_{j-1})$ using the same recalled sequences. If the correlation between the two sets of weights is not far above chance, or if removing the attention mechanism does not abolish backward contiguity and lower recall at small hidden dimensions, the claimed mechanistic equivalence is refuted.

Watch

Extended reading notes

Core claim

The central claim is a precise equivalence between two equations: the expected context reinstatement in CMR, $E[\alpha_j^{\mathrm{CMR}}] = \sum_i \frac{\exp(k c_{i-1}^\top c_{j-1})}{\sum_{i'} \exp(k c_{i'-1}^\top c_{j-1})} c_{i-1}$, and the attention context vector in the seq2seq model, $\alpha_j^{\mathrm{RNN}} = \sum_i \frac{\exp(h_i^\top h_j)}{\sum_{i'} \exp(h_{i'}^\top h_j)} h_i$. The paper derives this by aligning the decoder hidden state $h_j$ with the recall context $c_j$ and the encoder hidden states $h_i$ with the encoding contexts $c_{i-1}$, so that the two softmax distributions coincide. It then reports that the trained seq2seq model predicts individual free-recall behavior with lower root-mean-square error than CMR on the serial position curve, first-recall probability, and conditional response probability; that a reinforcement-learned model recovers the rational-CMR policy of starting at the first list item and recalling forward; and that ablating attention reproduces the lost backward contiguity and reduced recall seen in medial temporal lobe amnesia. The paper presents the mapping as a mechanistic correspondence, not just a metaphor, and uses the model to make novel predictions about working memory and episodic memory contributions.

Load-bearing premise

The whole argument rests on the claim that the decoder's internal state is literally the same kind of object as CMR's recall context, and the encoder's internal states are the encoding contexts; the paper assumes this identification rather than deriving it from the network update equations. If the trained network's internal states do not line up with CMR's contexts, attention and context reinstatement are only loosely analogous.

Editorial extensions

If this is right

A neural machine translation model can serve as an interpretable cognitive model of free recall: it predicts individual serial position curves, first-recall probabilities, and inter-item transition probabilities, and in the paper's fits does so with smaller error than CMR.
The independent convergence of translation-optimized architectures on context reinstatement is evidence that CMR's context architecture is functionally adaptive, not an arbitrary descriptive choice.
A reinforcement-learned seq2seq model recovers the optimal human free-recall policy—initiate recall at the first list item and continue forward—matching the optimal policy derived from CMR.
Hidden state size behaves like working memory capacity: smaller hidden states force the model to lean on attention (episodic context reinstatement) to achieve primacy, while larger hidden states can hold the list start in memory.
Ablating the attention mechanism reproduces the signature of medial temporal lobe amnesia—reduced recall and loss of backward contiguity—and the deficit disappears when hidden state size is large.

Reading between the lines

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

An extension the paper does not make: if the equivalence is real, then in transformer-based language models the same function should be carried by specific attention heads whose weights implement context reinstatement; the paper gestures at this link through recent work on induction heads, but does not establish it.
A testable extension: after fitting a seq2seq model to a subject, the fitted model's attention weights could be compared with that subject's trial-by-trial recall behavior; a close match would strengthen the mechanistic reading, while a mismatch would separate the two models.
The rational-convergence argument suggests a comparative experiment: train the same architecture on a non-translation sequence task with similar surface structure, and ask whether context reinstatement still emerges; if it does, the adaptive argument is about sequence prediction generally, not translation specifically.
The working-memory finding implies a human prediction: individuals with lower working memory capacity should show greater dependence on medial temporal lobe context reinstatement during free recall; the paper's simulations support this, but the human neural prediction is untested.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague — the thing to know: this paper has a genuinely interesting equation-level mapping between Luong attention and CMR context reinstatement, plus a set of behavioral simulations that go beyond the usual analogy. But the central derivation, as printed, has an index mismatch, and the stress-test note is correct on this point.

Eq. 22 defines E[α^CMR_j] using c_{j-1} as the probe context, while Eq. 23 defines α^RNN_j using h_j, the decoder state after consuming the embedding of f_{j-1}. The paper's own Eq. 17 and Figure 2 identify h_j with c_j. So the claimed equivalence E[α^CMR_j]≈α^RNN_j requires h_j to be both c_j and c_{j-1}. The fix is straightforward: shift the CMR index by one and use c_j as the probe, giving E[α^CMR_{j+1}] matching α^RNN_j. That repairability matters, because it means the core idea likely survives, but the proof as written is not sound.

What the paper does well: the mapping is the first detailed Luong-attention-to-CMR derivation I know, and the RL, working-memory-dimension, and attention-ablation analyses are real extensions. The individual-subject fits to PEERS are carefully done, with proper train/test splits and statistical tests. The authors also cite the prior transformer-to-CMR work (Li et al., Mistry et al.) rather than ignoring it. The hidden-dimension-as-working-memory story in Figure 5 is a nice addition that CMR cannot easily make.

Soft spots beyond the index issue: the implemented retrieval rule in Methods is not Luong dot-product attention. The model uses cosine similarity against an episodic memory table keyed by encoder hidden states, with GloVe embeddings as values. That differs from the architecture in the derivation, so the implemented model is one step removed from the equivalence claim. Also, no code and no reported CMR parameter values, which makes the quantitative comparisons hard to audit. And the RL reward is the same free recall reward used to define rational-CMR's optimal policy, so convergence to rational-CMR behavior is partly circular. None of these are fatal, but they should be addressed.

My bottom line: this is a serious, useful paper with a load-bearing technical flaw that is likely repairable. It deserves a rigorous peer review, not a desk rejection. If I were handling it, I would ask for the index fix, a clear statement of the retrieval-rule deviation, and code/data release before accepting. I would bring it to a reading group; there is enough here to argue about.

Referee Report

3 major / 4 minor

Summary. The paper argues that an RNN-based sequence-to-sequence model with Luong attention mechanistically corresponds to the Context Maintenance and Retrieval (CMR) model of human free recall. It develops a component-by-component mapping: encoder hidden states correspond to CMR encoding contexts, decoder hidden states correspond to recall contexts, and the attention context vector corresponds to CMR's context reinstatement mechanism. The authors then implement a seq2seq model as a cognitive model of free recall, fit it to individual PEERS subjects, compare it with CMR, train an RL version and compare its optimal behavior with rational-CMR, interpret hidden-state dimension as working-memory capacity, and ablate attention to model MTL amnesia. They conclude that the convergence between NMT architectures and CMR supports a rational account of context-based memory architecture and provides an interpretable, flexible model of memory search.

Significance. If the central equivalence were established, the paper would be a valuable cross-disciplinary contribution: it would provide an interpretable neural model of free recall with learning dynamics, and it would offer an evolutionary/rational argument for why the human memory system uses slowly drifting context and context reinstatement. The empirical scope is a genuine strength: per-subject fits on a large public dataset, RL optimization with convergence checks, a working-memory-capacity manipulation through hidden dimension, and an attention ablation tied to amnesia data. The paper is also transparent enough that its mathematical assumptions can be checked exactly, which is how the load-bearing problem in Eqs. 22-23 becomes visible. As printed, however, the headline equivalence is asserted under an index identification that contradicts the paper's own Eq. 17, and the implemented model uses a different retrieval rule from the Luong attention analyzed in the mapping. Both defects appear repairable, but the current version does not yet establish the central mechanistic claim.

major comments (3)

[Mathematical mapping, Eqs. 22-23] The claimed equivalence E[α^CMR_j] ≈ α^RNN_j does not follow from the displayed equations. Eq. 22 probes with c_{j-1}, the context before the just-recalled item is incorporated, whereas Eq. 23's probe h_j is the decoder state after the embedding x_j of f_{j-1} has been consumed (Eq. 2); under Eqs. 17-18 and Figure 2, the authors themselves identify this h_j with c_j, not c_{j-1}. The two softmax kernels can be aligned only by either redefining h_j as the decoder state before consuming f_{j-1} or shifting the CMR object to E[α^CMR_{j+1}] = Σ_i exp(k c_{i-1}^T c_j)/Σ_{i'} exp(k c_{i'-1}^T c_j) c_{i-1}. As written, the central equivalence is asserted rather than derived, although the mismatch appears repairable by a consistent index shift.
[Methods, 'Additional Model Details'] The behavioral implementation does not use the Luong dot-product attention that the mapping analyzes. After forming \hat h_j via Eq. 5, the model retrieves r_j = Σ_i sim(\hat h_j, h_i) x_i, where sim is cosine similarity and the stored values are GloVe embeddings in an episodic-memory table keyed by encoder hidden states, and then maps r_j through an inverse-embedding MLP to an item. This is a different attention-like operation from α^RNN_j = Σ_i softmax(h_j^T h_i) h_i in Eq. 23: it uses a different query (\hat h_j rather than h_j), a different similarity function, and different values (word embeddings rather than encoder states). Consequently the per-subject fits, RL optimization, hidden-dimension analyses, and ablation results validate a distinct retrieval model, and the paper has not shown empirically that the Luong-attention mechanism itself behaves like CMR's context reinstatement. The authors should either implement the mapped Luong mechanism in the model or explicitly derive the mapping for the cosine-similarity episodic-memory retrieval rule and separate the two claims.
[Discussion, 'A rational explanation of architectural assumptions'] The argument that the history of neural machine translation constitutes an effective rational search over architectures is not operationalized. The paper asserts that the NMT field 'has effectively conducted this analysis' over the space of memory-search architectures, but no search space, optimality criterion, or candidate comparison set is specified, so the claim that CMR's context-based architecture is adaptive is not a testable conclusion of this study. This does not undermine the mapping itself, but in a revision the authors should either provide a formal framework for the rational-architecture claim or present it explicitly as a motivating hypothesis rather than a demonstrated implication.

minor comments (4)

[Figure 2 and mapping text] The sentence immediately before Eq. 22 parenthetically identifies the probe as 'h_j or c_{j-1}', which is the same index conflict as in Eq. 17; please standardize the indexing in Figure 2 and in the surrounding text once the off-by-one issue is resolved.
[Results, model comparison statistics] Three Wilcoxon signed-rank tests on the same 171 subjects are reported without correction for multiple comparisons; please state whether the very small p-values survive a Bonferroni or false-discovery-rate correction.
[General reproducibility] The text refers to Supplementary Materials S2 and S3, but the arXiv submission does not include them, and no code or data availability statement is provided; please include the supplement or summarize the robustness checks in the main text, and state where the model implementation and processing scripts will be archived.
[Working memory capacity analogy] The mapping of hidden-dimension size to working memory capacity is acknowledged as only a relative measure; to strengthen this analogy, please relate it to quantitative models of working memory capacity or present it as a purely computational parameter.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the seq2seq-to-CMR mapping is an analytic equivalence drawn from the models' equations, and the behavioral and optimal-policy comparisons are genuine held-out or de-novo training results rather than fitted inputs reused as predictions.

full rationale

The central derivation (Eqs. 22-23) compares the CMR expected reactivated context, E[α^CMR_j] = Σ_i [exp(k c_{i-1}^T c_{j-1}) / Σ_{i'} exp(k c_{i'-1}^T c_{j-1})] c_{i-1}, with the seq2seq attention context, α^RNN_j = Σ_i [exp(h_i^T h_j) / Σ_{i'} exp(h_{i'}^T h_j)] h_i. The claimed equivalence is an analytic identification between two independently defined model families; it does not fit a parameter and then rename that fit as a prediction. The individual-subject comparisons are genuine predictions: the seq2seq model is trained on a training split and evaluated on a held-out test split, with CMR separately fit as a baseline. The optimized-behavior comparison in Figure 4 uses the same recall reward for the seq2seq RL agent and for rational-CMR, but that shared objective is an external task definition, not a fitted input, and the seq2seq agent is trained de novo, so its convergence to forward recall is an emergent simulation result rather than a circular consequence. The citation to Zhang et al. (2023) for rational-CMR is a same-group citation, but it is used as a published benchmark, and the current seq2seq experiments are self-contained; no load-bearing premise is justified only by that citation. One non-circular correctness concern should be noted: Eq. 22 probes with c_{j-1}, while Eq. 23 probes with h_j after the decoder has consumed f_{j-1}, and Eq. 17 aligns h_j with c_j, so the indexing of the advertised equivalence is not actually derived as written. That is a soundness gap, not a circular reduction; if repaired by shifting the CMR index by one step, the central claim would still be an independent derivation rather than an identity imposed by definition.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The mapping claim borrows CMR's context-update equations and adds an identification between hidden states and contexts; the behavioral claims add per-subject fitted parameters, hand-set RL rewards, and hand-chosen hidden dimension sizes. The paper's contribution is real but rests on several modeling choices that are not independently constrained.

free parameters (4)

CMR parameters (ρ, β, ρ', β', γ_FC, k) = not reported; per-subject Bayesian optimization
The claim that seq2seq fits human recall as well as CMR depends on these fitted values, but their ranges and final values are not given.
Hidden dimension size D = 32, 64, 128
Chosen by hand to represent working memory capacity; no independent criterion links D to human capacity.
RL reward penalties = +1 correct, -1 incorrect, -0.5 repeat; discount 0.99
Chosen by hand; the optimal policy and its comparison to rational-CMR depend on this reward structure.
Sinkhorn temperature τ = 1.0
Controls the sharpness of the set-prediction pre-training loss; not central to the main claim but affects learned representations.

assumptions (5)

ad hoc to paper Seq2seq decoder hidden states h_j can be identified with CMR recall contexts c_j, and encoder hidden states h_i with CMR encoding contexts c_{i-1}.
This identification is asserted in the derivation of Eqs. 22-23 and is never derived from the GRU update equations; it is the load-bearing assumption for the claimed equivalence.
domain assumption GloVe word embeddings serve as CMR's pre-experimental item-to-context associations MFC_pre.
Methods use GloVe vectors as the item embeddings x_i; the mapping to CMR requires treating them as pre-experimental context associations.
ad hoc to paper The field of neural machine translation has effectively searched over architectures in a way that approximates rational architecture optimization.
The adaptive-purpose argument depends on treating NMT's development as a proxy for exhaustive architecture search; this is an interpretive claim, not a measured fact.
domain assumption CMR is a valid descriptive model of human free recall.
The paper uses CMR as the reference cognitive model without quantitatively testing alternative memory models such as SAM.
ad hoc to paper RNN hidden dimension size is a relative measure of working memory capacity.
Used to interpret Figure 5 and the amnesia ablation; no independent validation of this mapping is given.

invented entities (1)

Episodic memory table keyed by encoder hidden states with GloVe embeddings as values
purpose: Implements CMR-like context-to-item retrieval in the seq2seq model and constrains recalls to studied items
This retrieval table is added in Methods and is not a standard component of the Luong attention seq2seq model; it has no falsifiable handle outside this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sequence-to-Sequence Models with Attention Mechanistically Map to the Architecture of Human Memory Search." pith.science (2026). https://pith.science/paper/PKAULZHK

@misc{pith2026250617424,
  author       = {Pith},
  title        = {Pith review of: Sequence-to-Sequence Models with Attention Mechanistically Map to the Architecture of Human Memory Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PKAULZHK}},
  note         = {Machine review of arXiv:2506.17424}
}

read the original abstract

Past work has long recognized the important role of context in guiding how humans search their memory. While context-based memory models can explain many memory phenomena, it remains unclear why humans develop such architectures over possible alternatives in the first place. In this work, we demonstrate that foundational architectures in neural machine translation -- specifically, recurrent neural network (RNN)-based sequence-to-sequence models with attention -- exhibit mechanisms that directly correspond to those specified in the Context Maintenance and Retrieval (CMR) model of human memory. Since neural machine translation models have evolved to optimize task performance, their convergence with human memory models provides a deeper understanding of the functional role of context in human memory, as well as presenting new ways to model human memory. Leveraging this convergence, we implement a neural machine translation model as a cognitive model of human memory search that is both interpretable and capable of capturing complex dynamics of learning. We show that our model accounts for both averaged and optimal human behavioral patterns as effectively as context-based memory models. Further, we demonstrate additional strengths of the proposed model by evaluating how memory search performance emerges from the interaction of different model components.

Figures

Figures reproduced from arXiv: 2506.17424 by the authors.

**Figure 1.** Illustrating the parallels between neural machine translation and human memory search models. (A) A seq2seq model with attention begins machine translation tasks by encoding each word of the original sequence into its hidden state, updated iteratively at each timestep, using its RNN encoder. In the decoding stage, the decoder RNN receives the final hidden state from the encoding stage and generates a word in the tar… view at source ↗

**Figure 2.** Detailed mapping between components in the seq2seq model with attention and those in the CMR Model. During the encoding phase, both the seq2seq model with attention (A) and the CMR model (C) maintain an internal state (hidden state hi or context vector ci) that combines input features of a presented item (fi) and the previous encoding state (hi−1 or ci−1) at each step i. In CMR, two fixed parameters control the mixi… view at source ↗

**Figure 3.** Behavioral patterns for individual subjects and the model predictions. (A–C) Behavioral patterns for the first 12 subjects out of 171 subjects, reproduced from Experiment 1 of the PEERS free recall dataset49, overlaid with behavioral patterns of the CMR model and the seq2seq model with attention, trained over a separate subset of the same individual subject data. The behavioral patterns include (A) the serial positi… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Optimized and intermediate training of the Seq2Seq model with attention The fully optimized seq2seq model with attention (128-Dim) shows behavioral patterns that closely align with those of the rational-CMR model across three sets of free recall patterns: serial positi…

**Figure 5.** Figure 5: The effect of working memory capacity (hidden dimension size) on optimal recall behavior (A–C) Heat maps illustrating the evolution of attention weights for seq2seq models with varying hidden dimension sizes across training epochs. Attention weights are averaged across…

**Figure 6.** Figure 6: Comparing optimized model behavior under ablation of the attention mechanism with recall behavior of medial temporal lobe (MTL) amnesia patients. (A–C) Behavioral patterns of MTL amnesia patients demonstrating memory deficits on a free recall task, reproduced from Palo…

Review history (2 revisions) →

Discussion (0). Continue with ORCID 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

Context-Aware Multi-Turn Visual-Textual Reasoning in LVLMs via Dynamic Memory and Adaptive Visual Guidance
cs.CV 2025-09 reject novelty 3.0 of 10

The proposed CAMVR framework is not supported by verifiable evidence, and the manuscript itself labels its experimental results as fabricated.

Reference graph

Works this paper leans on

85 extracted references · 58 canonical work pages · cited by 1 Pith paper

[1]

L., Steyvers, M

Griffiths, T. L., Steyvers, M. & Firl, A. Google and the mind: Predicting fluency with PageRank. Psychol. science 18, 1069–1076 (2007)

2007
[2]

& Norman, K

Lu, Q., Hasson, U. & Norman, K. A. A neural network model of when to retrieve and encode episodic memories. eLife 11, e74445, DOI: 10.7554/elife.74445 (2022)

work page doi:10.7554/elife.74445 2022
[3]

L., Norman, K

Callaway, F., Griffiths, T. L., Norman, K. A. & Zhang, Q. Optimal metacognitive control of memory recall. Psychol. Rev. (2023)

2023
[4]

& Hill, F

Lampinen, A., Chan, S., Banino, A. & Hill, F. Towards mental time travel: a hierarchical memory for reinforcement learning agents. Adv. Neural Inf. Process. Syst. 34, 28182–28195 (2021)

2021
[5]

M., Siegelmann, H

Van de Ven, G. M., Siegelmann, H. T. & Tolias, A. S. Brain-inspired replay for continual learning with artificial neural networks. Nat. communications 11, 4069 (2020)

2020
[6]

Cho, K. et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014). 20/23

arXiv 2014
[7]

Sutskever, I., Vinyals, O. & Le, Q. V . Sequence to sequence learning with neural networks. Adv. neural information processing systems 27 (2014)

2014
[8]

Luong, M.-T. et al. Effective approaches to attention-based neural machine translation. Conf. on Empir. Methods Nat. Lang. Process. DOI: 10.18653/v1/d15-1166 (2015)

doi:10.18653/v1/d15-1166 2015

Show all 85 references

[9]

& Bengio, Y

Bahdanau, D., Cho, K. & Bengio, Y . Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 (2014)

2014 arXiv
[10]

Attention is all you need

Vaswani, A. Attention is all you need. arXiv preprint arXiv:1706.03762 (2017)

2017 arXiv
[11]

Bert: Pre-training of deep bidirectional transformers for language understanding

Devlin, J. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

2018 arXiv
[12]

Brown, T. B. Language models are few-shot learners. arXiv preprint arXiv:2005.14165 (2020)

2020 arXiv
[13]

Raffel, C. et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. machine learning research 21, 1–67 (2020)

2020
[14]

Anderson, J. R. & Bower, G. H. Recognition and retrieval processes in free recall. Psychol. Rev. 79, 97–123 (1972)

1972
[15]

A multicomponent theory of the memory trace

Bower, G. A multicomponent theory of the memory trace. Psychol. Learn. Motiv. 1, 229–325 (1967)

1967
[16]

Estes, W. K. Statistical theory of spontaneous recovery and regression. Psychol. Rev. 62, 145–154 (1955)

1955
[17]

Howard, M. W. & Kahana, M. J. A distributed representation of temporal context. J. Math. Psychol. 46, 269–299 (2002)

2002
[18]

Murdock, B. B. Context and mediators in a theory of distributed associative memory (TODAM2). Psychol. Rev. 104, 839 (1997)

1997
[19]

M., Norman, K

Polyn, S. M., Norman, K. A. & Kahana, M. J. A context maintenance and retrieval model of organizational processes in free recall. Psychol. Rev. 116, 129–156 (2009)

2009
[20]

J., Polyn, S

Lohnas, L. J., Polyn, S. M. & Kahana, M. J. Expanding the scope of memory search: Modeling intralist and interlist effects in free recall. Psychol. Rev. 122, 337–363 (2015)

2015
[21]

A., Norman, K

Cornell, C. A., Norman, K. A., Griffiths, T. L. & Zhang, Q. Improving memory search through model-based cue selection. Psychol. Sci. 35, 55–71 (2024)

2024
[22]

Logan, G. D. & Cox, G. E. Serial memory: Putting chains and position codes in context. Psychol. Rev. 128, 1197–1205 (2021)

2021
[23]

Lohnas, L. J. A retrieved context model of serial recall and free recall. Comput. Brain & Behav. 8, 1–35 (2025)

2025
[24]

& Bhatia, S

Richie, R., Aka, A. & Bhatia, S. Free association in a neural network. Psychol. Rev. 130, 1360 (2023)

2023
[25]

Angne, H., Cornell, C. A. & Zhang, Q. A context-based model of collaborative inhibition during memory search. Sci. Reports 14, 27645 (2024)

2024
[26]

Zhou, Z., Kahana, M. J. & Schapiro, A. C. A unifying account of replay as context-driven memory reactivation. eLife 13 (2024)

2024
[27]

A., Niv, Y

Rouhani, N., Norman, K. A., Niv, Y . & Bornstein, A. M. Reward prediction errors create event boundaries in memory. Cognition 203, 104269, DOI: 10.1016/j.cognition.2020.104269 (2020)

2020
[28]

Y ., Talmi, D., Daw, N

Zhou, C. Y ., Talmi, D., Daw, N. D. & Mattar, M. G. Episodic retrieval for model-based evaluation in sequential decision tasks. Psychol. Rev. 132, 18–49 (2025)

2025
[29]

Anderson, J. R. & Milson, R. Human memory: An adaptive perspective. Psychol. Rev. 96, 703–719 (1989)

1989
[30]

Anderson, J. R. The adaptive character of thought (Psychology Press, New York, 1990)

1990
[31]

Zhang, Q., Griffiths, T. L. & Norman, K. A. Optimal policies for free recall. Psychol. Rev. 130, 1104 (2023)

2023
[32]

Murdock, B. B. The serial position effect of free recall. J. Exp. Psychol. 64, 482–488, DOI: 10.1037/h0045106 (1962)

1962 doi
[33]

Kahana, M. J. Associative retrieval processes in free recall. Mem. & Cogn. 24, 103–109, DOI: 10.3758/bf03197276 (1996)

1996 doi
[34]

Howard, M. W. & Kahana, M. J. Contextual variability and serial position effects in free recall. J. Exp. Psychol. Learn. Mem. Cogn. 25, 923 (1999)

1999
[35]

Reed, J. M. & Squire, L. R. Retrograde amnesia for facts and events: findings from four new cases. J. Neurosci. 18, 3943–3954 (1998). 21/23

1998
[36]

J., Di Lascio, J

Palombo, D. J., Di Lascio, J. M., Howard, M. W. & Verfaellie, M. Medial temporal lobe amnesia is associated with a deficit in recovering temporal context. J. cognitive neuroscience 31, 236–248 (2019)

2019
[37]

W., Fotedar, M

Howard, M. W., Fotedar, M. S., Datey, A. V . & Hasselmo, M. E. The temporal context model in spatial navigation and relational learning: toward a common explanation of medial temporal lobe function across domains. Psychol. Rev. 112, 75 (2005)

2005
[38]

Scoville, W. B. & Milner, B. Loss of recent memory after bilateral hippocampal lesions. J. neurology, neurosurgery, psychiatry 20, 11 (1957)

1957
[39]

& Church, K

Wang, H., Wu, H., He, Z., Huang, L. & Church, K. W. Progress in machine translation. Engineering 18, 143–153 (2022)

2022
[40]

F., Della Pietra, S

Brown, P. F., Della Pietra, S. A., Della Pietra, V . J. & Mercer, R. L. The mathematics of statistical machine translation: Parameter estimation. Comput. linguistics 19, 263–311 (1993)

1993
[41]

Koehn, P., Och, F. J. & Marcu, D. Statistical phrase-based translation. In 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology (HLT-NAACL 2003), 48–54 (Association for Computational Linguistics, 2003)

2003
[42]

Hinton, G. et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine 29, 82–97 (2012)

2012
[43]

& Hinton, G

Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. Adv. neural information processing systems 25 (2012)

2012
[44]

Elman, J. L. Finding structure in time. Cogn. science 14, 179–211 (1990)

1990
[45]

Bower, G. H. Stimulus-sampling theory of encoding variability. Coding processes human memory 3, 85–123 (1972)

1972
[46]

Kahana, M. J. Computational models of memory search. Annu. Rev. Psychol. 71, 107–138 (2020)

2020
[47]

Memory and consciousness

Tulving, E. Memory and consciousness. Can. Psychol. canadienne 26, 1 (1985)

1985
[48]

Pennington, J. et al. Glove: Global vectors for word representation. Conf. on Empir. Methods Nat. Lang. Process. DOI: 10.3115/v1/d14-1162 (2014)

2014 doi
[49]

Kahana, M. J. et al. The Penn electrophysiology of encoding and retrieval study. J. Exp. Psychol. Learn. Mem. Cogn. (2022)

2022
[50]

Shannon, C. E. A mathematical theory of communication. The Bell system technical journal 27, 379–423 (1948)

1948
[51]

P., Stokes, M

Piwek, E. P., Stokes, M. G. & Summerfield, C. A recurrent neural network model of prefrontal brain activity during a working memory task. PLoS Comput. Biol. 19, e1011555 (2023)

2023
[52]

Natural constraints explain working memory capacity limitations in sensory-cognitive models

Xie, Y .et al. Natural constraints explain working memory capacity limitations in sensory-cognitive models. bioRxiv 2023–03 (2023)

2023
[53]

& Schulz, E

Binz, M. & Schulz, E. Using cognitive psychology to understand gpt-3. Proc. Natl. Acad. Sci. 120, e2218523120 (2023)

2023
[54]

C., Bourgin, D

Peterson, J. C., Bourgin, D. D., Agrawal, M., Reichman, D. & Griffiths, T. L. Using large-scale experiments and machine learning to discover theories of human decision-making. Science 372, 1209–1214 (2021)

2021
[55]

Plonsky, O. et al. Predicting human decisions with behavioral theories and machine learning. arXiv preprint arXiv:1904.06866 (2019)

2019 arXiv
[56]

Kuperwajs, I., Schütt, H. H. & Ma, W. J. Using deep neural networks as a guide for modeling human planning. Sci. reports 13, 20269 (2023)

2023
[57]

Agrawal, M., Peterson, J. C. & Griffiths, T. L. Scaling up psychology via scientific regret minimization. Proc. Natl. Acad. Sci. 117, 8825–8835 (2020)

2020
[58]

& Mullainathan, S

Fudenberg, D., Kleinberg, J., Liang, A. & Mullainathan, S. Measuring the completeness of theories. arXiv preprint arXiv:1910.07022 (2019)

2019 arXiv
[59]

& Naecker, J

Peysakhovich, A. & Naecker, J. Using methods from machine learning to evaluate behavioral models of choice under risk and ambiguity. J. Econ. Behav. & Organ. 133, 373–384 (2017)

2017
[60]

Ho, M. K. & Griffiths, T. L. Cognitive science as a source of forward and inverse models of human decisions for robotics and control. Annu. Rev. Control. Robotics, Auton. Syst. 5, 33–53 (2022)

2022
[61]

Liu, N. F. et al. Lost in the middle: How language models use long contexts. arXiv preprint arXiv:2307.03172 (2023)

2023 arXiv
[62]

& V osoughi, S

Guo, X. & V osoughi, S. Serial position effects of large language models.arXiv preprint arXiv:2406.15981 (2024). 22/23

2024 arXiv
[63]

& Mattar, M

Li, J.-A., Zhou, C., Benna, M. & Mattar, M. G. Linking in-context learning in transformers to human episodic memory. Adv. Neural Inf. Process. Syst. 37, 6180–6212 (2024)

2024
[64]

M., Bajaj, A., Aggarwal, Y ., Maini, S

Mistry, D. M., Bajaj, A., Aggarwal, Y ., Maini, S. S. & Tiganj, Z. Emergence of episodic memory in transformers: Characterizing changes in temporal structure of attention scores during training. arXiv preprint arXiv:2502.06902 (2025)

2025 arXiv
[65]

Fountas, Z. et al. Human-like episodic memory for infinite context llms. arXiv preprint arXiv:2407.09450 (2024)

2024
[66]

& Cunitz, A

Glanzer, M. & Cunitz, A. R. Two storage mechanisms in free recall. J. V erbal Learn. V erbal Behav.5, 351–360 (1966)

1966
[67]

Bjork, R. A. & Whitten, W. B. Recency-sensitive retrieval processes in long-term free recall⋆. Cogn. Psychol. 6, 173–189 (1974)

1974
[68]

Raaijmakers, J. G. & Shiffrin, R. M. Search of associative memory. Psychol. Rev. 88, 93–134 (1981)

1981
[69]

B., Conway, A

Morrison, A. B., Conway, A. R. & Chein, J. M. Primacy and recency effects as indices of the focus of attention. Front. human neuroscience 8, 6 (2014)

2014
[70]

X., Dam, C

Ranganath, C., Cohen, M. X., Dam, C. & D’Esposito, M. Inferior temporal, prefrontal, and hippocampal contributions to visual working memory maintenance and associative memory retrieval. J. Neurosci. 24, 3917–3925 (2004)

2004
[71]

R., Page, K., Moore, K

Olson, I. R., Page, K., Moore, K. S., Chatterjee, A. & Verfaellie, M. Working memory for conjunctions relies on the medial temporal lobe. J. Neurosci. 26, 4596–4601 (2006)

2006
[72]

& Squire, L

Jeneson, A. & Squire, L. R. Working memory, long-term memory, and medial temporal lobe function. Learn. & memory 19, 15–25 (2012)

2012
[73]

M., Forstmann, B

Turner, B. M., Forstmann, B. U., Love, B. C., Palmeri, T. J. & Van Maanen, L. Approaches to analysis in model-based cognitive neuroscience. J. Math. Psychol. 76, 65–79 (2017)

2017
[74]

Vision: A computational investigation into the human representation and processing of visual information (MIT press, 2010)

Marr, D. Vision: A computational investigation into the human representation and processing of visual information (MIT press, 2010)

2010
[75]

L., Lieder, F

Griffiths, T. L., Lieder, F. & Goodman, N. D. Rational use of cognitive resources: Levels of analysis between the computational and the algorithmic. Top. Cogn. Sci. 7, 217–229 (2015)

2015
[76]

& Zhang, Q

Xu, Z., Hemmer, P. & Zhang, Q. Towards a generalized bayesian model of reconstructive memory. Comput. Brain & Behav. (2024)

2024
[77]

Huttenlocher, J., Hedges, L. V . & Vevea, J. L. Why do categories affect stimulus judgment?J. experimental psychology: Gen. 129, 220 (2000)

2000
[78]

Simon, H. A. Rationality as process and as product of thought. The Am. Econ. Rev. 68, 1–16 (1978)

1978
[79]

Howes, A., Lewis, R. L. & Vera, A. Rational adaptation under task and processing constraints: implications for testing theories of cognition and action. Psychol. review 116, 717 (2009)

2009
[80]

Bayesian Optimization: Open source constrained global optimization tool for Python (2014–)

Nogueira, F. Bayesian Optimization: Open source constrained global optimization tool for Python (2014–)

2014
[81]

Bahdanau, D. et al. An actor-critic algorithm for sequence prediction. ArXiv abs/1607.07086 (2016)

2016 arXiv
[82]

& Zaremba, W

Ranzato, M., Chopra, S., Auli, M. & Zaremba, W. Sequence level training with recurrent neural networks. CoRR abs/1511.06732 (2015)

2015 arXiv
[83]

Sinkhorn distances: Lightspeed computation of optimal transport

Cuturi, M. Sinkhorn distances: Lightspeed computation of optimal transport. In Neural Information Processing Systems (2013)

2013
[84]

E., Belanger, D., Linderman, S

Mena, G. E., Belanger, D., Linderman, S. W. & Snoek, J. Learning latent permutations with gumbel-sinkhorn networks. ArXiv abs/1802.08665 (2018)

2018 arXiv
[85]

& Klimov, O

Schulman, J., Wolski, F., Dhariwal, P., Radford, A. & Klimov, O. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017). 23/23

2017 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

L., Steyvers, M

Griffiths, T. L., Steyvers, M. & Firl, A. Google and the mind: Predicting fluency with PageRank. Psychol. science 18, 1069–1076 (2007)

2007

[2] [2]

& Norman, K

Lu, Q., Hasson, U. & Norman, K. A. A neural network model of when to retrieve and encode episodic memories. eLife 11, e74445, DOI: 10.7554/elife.74445 (2022)

work page doi:10.7554/elife.74445 2022

[3] [3]

L., Norman, K

Callaway, F., Griffiths, T. L., Norman, K. A. & Zhang, Q. Optimal metacognitive control of memory recall. Psychol. Rev. (2023)

2023

[4] [4]

& Hill, F

Lampinen, A., Chan, S., Banino, A. & Hill, F. Towards mental time travel: a hierarchical memory for reinforcement learning agents. Adv. Neural Inf. Process. Syst. 34, 28182–28195 (2021)

2021

[5] [5]

M., Siegelmann, H

Van de Ven, G. M., Siegelmann, H. T. & Tolias, A. S. Brain-inspired replay for continual learning with artificial neural networks. Nat. communications 11, 4069 (2020)

2020

[6] [6]

Cho, K. et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014). 20/23

arXiv 2014

[7] [7]

Sutskever, I., Vinyals, O. & Le, Q. V . Sequence to sequence learning with neural networks. Adv. neural information processing systems 27 (2014)

2014

[8] [8]

Luong, M.-T. et al. Effective approaches to attention-based neural machine translation. Conf. on Empir. Methods Nat. Lang. Process. DOI: 10.18653/v1/d15-1166 (2015)

doi:10.18653/v1/d15-1166 2015