Exponential Family Attention

Kevin Christian Wibisono; Yixin Wang

REVIEW 4 major objections 5 minor 30 references

Exponential Family Attention

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Exponential family attention turns self-attention into a probabilistic model for mixed-type data.

desk verdict The central identifiability theorem is false—the proof cancels a vector it can't cancel—so the paper's theoretical foundation collapses, despite a reasonable model idea. read the letter →

arxiv 2501.16790 v1 pith:V34JZEA6 submitted 2025-01-28 stat.ML cs.LG

Kevin Christian Wibisono , Yixin Wang This is my paper · ORCID

classification stat.MLcs.LG

keywords attentionmechanismexponentialfamilieslatentfactormodelsprobabilisticmodelingrepresentationlearningself-attentionidentifiabilitygeneralizationbound

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 proposes exponential family attention (EFA), a probabilistic generative model that extends the transformer's self-attention mechanism beyond text to high-dimensional sequence, spatial, and spatiotemporal data with mixed discrete and continuous observations. EFA conditions each observation on all others through an attention-weighted latent factor model, so the relevance of each context point is itself learned rather than fixed. The paper argues that EFA is linearly identifiable under diversity assumptions, comes with an excess-loss generalization guarantee, and subsumes classical linear latent factor models as special cases while outperforming them on synthetic data, movie ratings, shopping baskets, and city temperatures. If correct, EFA provides a unified, theoretically supported attention-based framework for modeling mixed-type data.

What carries the argument

The central object is the EFA model itself: each observation is predicted from a masked context matrix whose columns combine center embeddings, context embeddings, value embeddings, and positional embeddings, passed through multi-head self-attention; the masked column is mapped to the natural parameter of an exponential family distribution (categorical for discrete tokens, Gaussian, Poisson, or ordinal for associated values). The identifiability argument rests on the diversity Assumptions 2–4, adapted from Roeder et al. (2021), which require the existence of small data sets making certain embedding-difference matrices invertible; the generalization argument follows the transformer model and covering-number estimate of Bai et al. (2023).

What would settle it

Construct two parameter settings of EFA on a real dataset (e.g., MovieLens) that yield identical conditional distributions, then check numerically whether the learned center/context embeddings are necessarily related by an invertible linear transformation; if a pair is found where they are not, Theorem 1's conclusion fails. Alternatively, a concrete Gaussian with a nonlinear reparametrization where equality of means holds without any invertible linear map between embeddings would refute the unqualified claim.

Watch

Extended reading notes

Core claim

The central claim is that self-attention can be turned into a general probabilistic model by letting the attention mechanism produce the natural parameters of an exponential family distribution for each observation given its context. Concretely, EFA models categorical observations with the standard softmax attention, and models associated continuous or discrete values through an exponential family whose natural parameter is derived from a transformer applied to concatenated category, value, and positional embeddings. The paper proves that under three diversity assumptions plus a structural decomposition assumption, the learned center and context embeddings are unique up to invertible linear transformations (Theorem 1), and that empirical risk minimization achieves a population excess loss of order $\tilde{O}(B_y^2 \sqrt{[(L(3MD^2 + 2DD') + K'(K + \tau))\iota + \log(1/\xi)]/F})$ (Theorem 2). It also shows constructively that linear latent factor models are a special case, so any improvement over them reflects the additional modeling capacity.

Load-bearing premise

The identifiability theorem rests on the unverified assumption that for any two parameter settings that give identical predictions, one can always find a small set of data points making certain embedding-difference matrices invertible (Assumptions 2–4); on real datasets this is not established.

Editorial extensions

If this is right

EFA gives a single probabilistic recipe for mixed-type sequential data, so models that previously required separate treatments for discrete and continuous observations can share one attention backbone.
Because EFA provably contains linear latent factor models as special cases, any empirical gap between EFA and such models measures the value of dynamic, context-dependent relevance weighting.
The linear identifiability result means the learned representations are unique up to linear transformations, making downstream analyses of embeddings (e.g., finding complements in shopping baskets) more meaningful.
The excess-loss guarantee quantifies how much training data is needed for the empirical risk minimizer to approach the best population predictor in the model class.

Reading between the lines

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

If the paper's claims hold, EFA could be applied to other mixed-type spatiotemporal and relational data, such as sensor networks or electronic health records, where discrete events and continuous measurements co-occur.
The identifiability assumptions are plausibly hard to verify on real data; a direct check on a real dataset (e.g., whether two fitted EFA solutions with equal predictive distributions have embeddings related by a linear map) would settle the practical reach of Theorem 1.
If the Gaussian counterexample with nonlinear reparametrization generalizes, then linear identifiability may require restricting the embedding maps to linear or injective forms even when Assumptions 2–4 hold.

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.

Desk Editor's Note

read the letter

The useful thing to know: this paper's central theoretical claim is wrong. The identifiability result (Theorem 1) relies on an invalid vector cancellation, and a simple two-dimensional Gaussian counterexample satisfies the paper's own Assumptions 1 and 4 while violating the claimed linear identifiability. The empirical evaluation is also weak: no error bars anywhere, and the temperature baseline's neighborhood size is selected on the test set.

What's genuinely new: EFA is a clean way to make exponential-family embedding contexts data-dependent via self-attention. The model formulation is sensible, and Propositions 1 and 2 correctly show that linear latent factor models are special cases. The paper is clearly written and engages honestly with the existing literature. Theorem 2's generalization bound is an application of Bai et al.'s machinery, and it appears to be adapted correctly.

The soft spots are not minor. In the proof of Theorem 1's θ2 part, from (N')ᵀL' = (N*)ᵀL* you cannot conclude (N')ᵀ = (N*)ᵀD⁻¹. The correct inference is only an orthogonality condition. The counterexample from the stress-test—K*(y) = (cos y, sin y)ᵀ, K'(y) = (cos y, sin²y + sin y)ᵀ, with L' = L* = (1,0)ᵀ—gives identical conditional distributions for all y, satisfies Assumption 4 at y = 0 and y = π/2, yet no invertible C maps K' to K*. So Definition 3 collapses, and Theorem 1 is false as stated. Separately, Proposition 3's proof defines W^V twice with different matrices, which is at best a typo and at worst an inconsistent construction. The experiments only compare against a simple linear FM, with no error bars; the temperature experiment picks the FM's k by test performance, which is a methodological red flag.

Who is this for? Someone working on probabilistic attention or exponential-family embeddings might find the model formulation suggestive, but the paper in its current form doesn't support its claims. It needs major revision: either fix or drop the identifiability theorem, and redo the empirical comparison with proper baselines and variance estimates.

I wouldn't cite it yet, but I'd send it to a serious referee. The model idea is worth discussing, and the flaw is exactly the kind a careful referee should catch. It's not a desk reject in terms of importance, but it requires substantial revision before acceptance.

Referee Report

4 major / 5 minor

Summary. The paper introduces exponential family attention (EFA), a probabilistic generative model that combines self-attention with exponential family conditional distributions to handle mixed-type sequence, spatial, and spatiotemporal data. The EFA likelihood factorizes into a categorical self-attention term for discrete observations and an exponential family term for associated continuous or discrete values. The authors prove (Theorem 1) that EFA is linearly identifiable under Assumptions 1–4, provide an excess-loss generalization guarantee (Theorem 2) for a simplified Gaussian spatiotemporal version, show that several linear latent factor models are special cases of EFA (Propositions 1–3), and report experiments on synthetic data, MovieLens, Instacart, and U.S. city temperatures where EFA outperforms linear latent factor models (FM).

Significance. If the results were correct, EFA would be a notable contribution: it unifies self-attention and exponential family embeddings, offers a principled probabilistic extension of attention to non-text data, and provides an identifiability guarantee plus a generalization bound. The paper also ships a public software repository and gives constructive proofs that EFA subsumes common linear latent factor models, which are useful didactic results. However, the central theoretical claim (Theorem 1) rests on an invalid proof step and is in fact false as stated, so the main contribution of the paper is not established. The empirical comparisons, while positive, are against only linear baselines and lack statistical rigor (no error bars, neighborhood size selected on the test set).

major comments (4)

[Appendix C, proof of Theorem 1] The proof of the θ2 identifiability part is invalid. From equality of conditional Gaussians, the proof obtains (N')ᵀ L' = (N*)ᵀ L*, then defines D = (N*(N')⁻¹)ᵀ and asserts that (N')ᵀ = (N*)ᵀ D⁻¹. This step cancels the vector L* without justification. The correct consequence is (Dᵀ K'(y) − K*(y)) · L* = 0 for every y, an orthogonality condition that does not imply K'(y) = C K*(y) for any invertible C. This gap is not cosmetic: take dⱼ=2, L*=L'=(1,0)ᵀ, K*(y)=(cos y, sin y)ᵀ, K'(y)=(cos y, sin²y+sin y)ᵀ. Then K'(y)ᵀ L' = cos y = K*(y)ᵀ L* for all y, and for y₁=0, y₂=π/2 the matrices N' and N* are both invertible, so Assumption 4 holds. Yet no invertible C satisfies K'=CK* because sin²y is not in the linear span of cos y and sin y. Thus the conclusion of Theorem 1 for θ2 is false, and Definition 3 collapses.
[Section 5.1.2, Assumptions 2–4] Assumptions 2–4 are asserted to be 'mild' and to 'hold almost surely' (Remark 13), but no verification is given for any real dataset or for the model classes used in the experiments. Since Theorem 1 depends on these diversity conditions, and since the proof itself is invalid, the identifiability claim is unsupported even under the stated assumptions.
[Section 6.4, Figure 6] The FM baseline's neighborhood size k is selected on the test set: the text reports the lowest test MSE among all FM configurations (k=5) and then compares EFA against that test-selected value. This invalidates the comparison as a rigorous evaluation of generalization, since the FM baseline is given an advantage of test-set tuning while EFA is not. The reported conclusion that 'EFA outperforms FM across all neighbor sizes' is still visible in Figure 6, but the headline comparison should be based on a validation-set-selected k.
[Section 6, Tables 1 and 2] All experimental results are reported without error bars, confidence intervals, or repeated-run variability. The claims that EFA 'consistently outperforms' FM rest on single point estimates; given the modest differences in several settings (e.g., Table 1 right panel: 3.444 vs 3.534; Table 3 lower rows), the possibility of run-to-run variance or initialization effects cannot be ruled out.

minor comments (5)

[Section 5.2.3, Theorem 2 statement] The bound contains a typographical ambiguity: the displayed term '43L(L(3M D2 + 2DD′) + K′(K + τ))ι' appears to be missing a square root or parentheses; the proof derives an expression with a square root, so the statement should be corrected to match.
[Section 5.2.1, loss formulation] The training instances are written as Z={(τ1,y1),...,(τI,yI)} but the subsequent line says 'Z f := {(τ f 1 , y f 1 ), · · ·, (τ f I , x f I )}'; the last coordinate should be y, not x.
[Section 3.2, Example 2] The notation y(i,f) is introduced but the text then writes 'the observation at time step f for the i-th time series' and later omits f; the notation is understandable but could be clarified by explicitly defining the full index set.
[Section 6.1, Table 1] The synthetic data experiment reports test MSE values of 4.519 and 1.033 for FM and EFA; given that the data generation includes Gaussian noise with variance 1, the large improvement is plausible but would be more convincing with a description of the number of synthetic users and the split used.
[Appendix H.4] The description of the EFA temperature model says each column of the matrix is passed through dense layers 'resulting in Yi' before applying self-attention; this is a minor notational inconsistency with the main text, where Yi denotes the raw input matrix.

Circularity Check

0 steps flagged · score 0.0 of 10

No part of the EFA derivation reduces to its own inputs by construction; the potential issues are correctness defects, not circularity.

full rationale

No step in the paper's derivation chain is circular under the specified criteria. The identifiability result splits into two parts: the θ1 half is imported from Roeder et al. (2021), an external theorem with assumptions restated as Assumptions 2 and 3, and the θ2 half is attempted from Assumptions 1 and 4 via Appendix C. Although the θ2 proof contains an unjustified vector cancellation, and the claim can fail under a two-dimensional Gaussian counterexample, this is a mathematical correctness defect rather than a circularity: Assumption 4's invertibility condition does not mention the transformation matrix C, so the conclusion K' = C K* is not contained in the assumption by construction. Theorem 2 is an ERM generalization bound obtained by applying Proposition B.4 of Bai et al. (2023), again independent external support rather than self-citation. The empirical section compares EFA against a linear latent factor special case on held-out test sets, so the reported improvements are not fitted parameters renamed as predictions. There are no load-bearing self-citations, no uniqueness theorems imported from the authors' own prior work, and no ansatz smuggled in via self-citation. The paper's main risks are the validity of Theorem 1 and the strength of its baselines, both of which are correctness or evaluation concerns, not circularity.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The EFA model does not introduce new particles, forces, or dimensions. It relies on learned embeddings and attention matrices, which are ordinary model parameters. The key load-bearing inputs are the assumptions listed above and the choice of baselines.

free parameters (2)

FM neighborhood size k (temperature experiment) = 5
Selected as the configuration with the lowest test MSE in Figure 6; this is a post-hoc test-set choice that biases the comparison.
EFA architecture hyperparameters = 32-dim embeddings; 2 layers, 2 heads; temperature uses 1, 2, or 4 layers
Hand-chosen model sizes are not swept; the reported improvements may depend on them, and no sensitivity analysis is given.

assumptions (5)

ad hoc to paper Assumption 1: J_{theta2}(y) = K_{theta2}(y)^T L_{theta2} for the Gaussian term.
This bilinear decomposition is introduced specifically for the identifiability theorem and restricts the EFA family.
ad hoc to paper Assumptions 2 to 4 (diversity): for any theta', theta* with equal conditional distributions, there exist small sets of data points making the matrices L', L*, M', M*, N', N* invertible.
These assumptions are not checked on any dataset and are strong enough to potentially preempt the identifiability conclusion.
domain assumption Assumption 5: |y_i| <= B_y and the loss is clipped; R is large enough not to change behavior.
Standard boundedness in transformer generalization proofs (Bai et al., 2023).
domain assumption Independence x_i independent of y_{1:i-1} given x_{1:i-1} in Eq. (2).
The factorized likelihood is an unverified modeling assumption for the real datasets.
domain assumption Training data in Theorem 2 are i.i.d. draws from P drawn from pi, with a separate time index per training instance.
Needed for the excess loss bound; the temperature data are a single fixed time series, so the i.i.d. assumption is a simplification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exponential Family Attention." pith.science (2026). https://pith.science/paper/V34JZEA6

@misc{pith2026250116790,
  author       = {Pith},
  title        = {Pith review of: Exponential Family Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V34JZEA6}},
  note         = {Machine review of arXiv:2501.16790}
}

read the original abstract

The self-attention mechanism is the backbone of the transformer neural network underlying most large language models. It can capture complex word patterns and long-range dependencies in natural language. This paper introduces exponential family attention (EFA), a probabilistic generative model that extends self-attention to handle high-dimensional sequence, spatial, or spatial-temporal data of mixed data types, including both discrete and continuous observations. The key idea of EFA is to model each observation conditional on all other existing observations, called the context, whose relevance is learned in a data-driven way via an attention-based latent factor model. In particular, unlike static latent embeddings, EFA uses the self-attention mechanism to capture dynamic interactions in the context, where the relevance of each context observations depends on other observations. We establish an identifiability result and provide a generalization guarantee on excess loss for EFA. Across real-world and synthetic data sets -- including U.S. city temperatures, Instacart shopping baskets, and MovieLens ratings -- we find that EFA consistently outperforms existing models in capturing complex latent structures and reconstructing held-out data.

Figures

Figures reproduced from arXiv: 2501.16790 by the authors.

**Figure 2.** An end-to-end illustration of the exponential family attention (EFA) model. The [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Unidirectional attention weights for two representative baskets. [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Per-item query, key, and value embeddings for each head of the first layer. [PITH_FULL_IMAGE:figures/full_fig_p027_4.png]

**Figure 5.** Figure 5: All cities included in the experiment in Section [PITH_FULL_IMAGE:figures/full_fig_p028_5.png]

**Figure 6.** Figure 6: EFA outperforms FM across all neighbor sizes ( [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

[1]

Bai, Y., F. Chen, H. Wang, C. Xiong, and S. Mei (2023). Transformers as statisticians: P rovable in-context learning with in-context algorithm selection. Neural Information Processing Systems\/ 36

work page 2023
[2]

Brown, T., B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. (2020). Language models are few-shot learners. Neural Information Processing Systems\/ 33 , 1877--1901

work page 2020
[3]

Celikkanat, A. and F. D. Malliaros (2020). Exponential family graph embeddings. AAAI Conference on Artificial Intelligence\/ 34\/ (04), 3357--3364

work page 2020
[4]

Chen, F., X. Liu, D. Proserpio, I. Troncoso, and F. Xiong (2020). Studying product competition using representation learning. ACM SIGIR Conference on Research and Development in Information Retrieval\/ , 1261--1268

work page 2020
[5]

Li, and S

Chen, Y., X. Li, and S. Zhang (2019). Joint maximum likelihood estimation for high-dimensional exploratory item factor analysis. Psychometrika\/ 84 , 124--146

work page 2019
[6]

Li, and S

Chen, Y., X. Li, and S. Zhang (2020). Structured latent factor analysis for large-scale data: I dentifiability, estimability, and their implications. Journal of the American Statistical Association\/ 115\/ (532), 1756--1770

work page 2020
[7]

Chang, K

Devlin, J., M.-W. Chang, K. Lee, and K. Toutanova (2019). BERT : P re-training of deep bidirectional transformers for language understanding. North American Chapter of the Association for Computational Linguistics\/

work page 2019
[8]

Gu, Y. and D. B. Dunson (2023). Bayesian pyramids: I dentifiable multilayer discrete latent structure models for discrete data. Journal of the Royal Statistical Society Series B: Statistical Methodology\/ 85\/ (2), 399--426

work page 2023

Show all 30 references

[9]

Gu, Y. and G. Xu (2023). A joint MLE approach to large-scale structured latent attribute analysis. Journal of the American Statistical Association\/ 118\/ (541), 746--760

2023
[10]

Celikkanat, A

Jagtap, S., A. Celikkanat, A. Piravre, F. Bidard, L. Duval, and F. D. Malliaros (2021). Multiomics data integration for gene regulatory network inference with exponential family embeddings. European Signal Processing Conference\/ , 1221--1225

2021
[11]

Kingma, D. P. and J. Ba (2015). Adam: A method for stochastic optimization. International Conference on Learning Representations\/

2015
[12]

Krstovski, K. and D. M. Blei (2018). Equation embeddings. arXiv preprint arXiv:1803.09123\/

2018 arXiv
[13]

Lin, K. Z., J. Lei, and K. Roeder (2021). Exponential-family embedding with application to cell developmental trajectories for single-cell RNA -seq data. Journal of the American Statistical Association\/ 116\/ (534), 457--470

2021
[14]

Lin, Z., M. Feng, C. N. d. Santos, M. Yu, B. Xiang, B. Zhou, and Y. Bengio (2017). A structured self-attentive sentence embedding. International Conference on Learning Representations\/

2017
[15]

Liu, L.-P. and D. M. Blei (2017). Zero-inflated exponential family embeddings. International Conference on Machine Learning\/ , 2140--2148

2017
[16]

Liu, Y., M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov (2019). Ro BERT a: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692\/

2019 arXiv
[17]

Peltonen, T

Lu, C., J. Peltonen, T. Nummenmaa, and J. Nummenmaa (2022). Nonparametric exponential family graph embeddings for multiple representation learning. Uncertainty in Artificial Intelligence\/ , 1275--1285

2022
[18]

Chen, and Y

Lyu, Z., L. Chen, and Y. Gu (2024). Degree-heterogeneous latent class analysis for high-dimensional discrete data. arXiv preprint arXiv:2402.18745\/

2024 arXiv
[19]

Sutskever, K

Mikolov, T., I. Sutskever, K. Chen, G. S. Corrado, and J. Dean (2013). Distributed representations of words and phrases and their compositionality. Neural Information Processing Systems\/ 26

2013
[20]

GPT -4 technical report

OpenAI (2023). GPT -4 technical report. arXiv preprint arXiv:2303.08774\/

2023 arXiv
[21]

Ouyang, J., C. Cui, K. M. Tan, and G. Xu (2024). Statistical inference for covariate-adjusted and interpretable generalized factor model with application to testing fairness. arXiv preprint arXiv:2404.16745\/

2024
[22]

Tackstrom, D

Parikh, A., O. Tackstrom, D. Das, and J. Uszkoreit (2016). A decomposable attention model for natural language inference. arXiv preprint arXiv:1606.01933\/

2016 arXiv
[23]

Metz, and D

Roeder, G., L. Metz, and D. Kingma (2021). On linear identifiability of learned representations. International Conference on Machine Learning\/ , 9030--9039

2021
[24]

Rudolph, M. and D. Blei (2018). Dynamic embeddings for language evolution. World Wide Web Conference\/ , 1003--1011

2018
[25]

Rudolph, M., F. Ruiz, S. Athey, and D. Blei (2017). Structured embedding models for grouped data. Neural Information Processing Systems\/ 30

2017
[26]

Rudolph, M., F. Ruiz, S. Mandt, and D. Blei (2016). Exponential family embeddings. Neural Information Processing Systems\/ 29

2016
[27]

Sun, J., Y. Chen, J. Liu, Z. Ying, and T. Xin (2016). Latent variable selection for multidimensional item response theory models via l1 regularization. Psychometrika\/ 81\/ (4), 921--939

2016
[28]

Shazeer, N

Vaswani, A., N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin (2017). Attention is all you need. Neural Information Processing Systems\/ 30

2017
[29]

Wainwright, M. J. (2019). High-dimensional statistics: A non-asymptotic viewpoint . Cambridge University Press

2019
[30]

Yu, M., Z. Yang, T. Zhao, M. Kolar, and Z. Wang (2018). Provable G aussian embedding with one observation. Neural Information Processing Systems\/ 31

2018

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Bai, Y., F. Chen, H. Wang, C. Xiong, and S. Mei (2023). Transformers as statisticians: P rovable in-context learning with in-context algorithm selection. Neural Information Processing Systems\/ 36

work page 2023

[2] [2]

Brown, T., B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. (2020). Language models are few-shot learners. Neural Information Processing Systems\/ 33 , 1877--1901

work page 2020

[3] [3]

Celikkanat, A. and F. D. Malliaros (2020). Exponential family graph embeddings. AAAI Conference on Artificial Intelligence\/ 34\/ (04), 3357--3364

work page 2020

[4] [4]

Chen, F., X. Liu, D. Proserpio, I. Troncoso, and F. Xiong (2020). Studying product competition using representation learning. ACM SIGIR Conference on Research and Development in Information Retrieval\/ , 1261--1268

work page 2020

[5] [5]

Li, and S

Chen, Y., X. Li, and S. Zhang (2019). Joint maximum likelihood estimation for high-dimensional exploratory item factor analysis. Psychometrika\/ 84 , 124--146

work page 2019

[6] [6]

Li, and S

Chen, Y., X. Li, and S. Zhang (2020). Structured latent factor analysis for large-scale data: I dentifiability, estimability, and their implications. Journal of the American Statistical Association\/ 115\/ (532), 1756--1770

work page 2020

[7] [7]

Chang, K

Devlin, J., M.-W. Chang, K. Lee, and K. Toutanova (2019). BERT : P re-training of deep bidirectional transformers for language understanding. North American Chapter of the Association for Computational Linguistics\/

work page 2019

[8] [8]

Gu, Y. and D. B. Dunson (2023). Bayesian pyramids: I dentifiable multilayer discrete latent structure models for discrete data. Journal of the Royal Statistical Society Series B: Statistical Methodology\/ 85\/ (2), 399--426

work page 2023