Pith. sign in

REVIEW 4 major objections 5 minor 95 references

When Do Neural Networks Learn World Models?

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A degree-minimizing multi-task learner recovers the true latent variables of its data up to sign flips and coordinate permutations, provided the proxy-task distribution is tilted toward low-complexity tasks.

desk verdict A serious Boolean-function theory result for when low-degree bias recovers latents, with one real caveat: the p1>0 assumption is more substantive than 'mild,' and the paper itself admits it. read the letter →

arxiv 2502.09297 v5 pith:GAXBHNQM submitted 2025-02-13 cs.LG

Tianren Zhang , Guanyu Chen , Feng Chen This is my paper
classification cs.LG MSC 68T0706E3068Q32
keywords worldmodelslatentvariablerecoveryBooleanfunctionsFourier-Walshtransformlow-degreebiasmulti-tasklearningbasiscompatibilityout-of-distributiongeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks a concrete version of a fuzzy question: when does a neural network actually learn the hidden variables that generate its training data, rather than a superficial shortcut? The authors model observed data as a Boolean function $x=\psi(z)$ of latent bits $z$, and model a learner's complexity bias as minimization of the realization degree—the sum of Fourier-Walsh degrees of the functions in a compositional solution. Their central result, Theorem 4.8, says that in a multi-task setting where proxy tasks are drawn from a degree-biased distribution, the degree-minimizing solution provably recovers $z$ up to negations and permutations, even when every task is a nonlinear function of the latents. This matters because the classical non-identifiability of nonlinear latent-variable models (Lemma 2.4, following Hyvärinen and Pajunen 1999) had suggested such recovery was impossible without explicit distributional assumptions. The paper also shows recovery is architecture-sensitive and that a recovered world model provably improves out-of-distribution generalization in a Hamming-ball setting.

What carries the argument

The load-bearing object is the realization degree, $\mathrm{ddeg}(f^{(1)}\circ\dots\circ f^{(q)})=\sum_i \deg(f^{(i)})$, computed using the Fourier-Walsh degree of each Boolean function—the size of the largest parity monomial $\chi_S(x)=\prod_{i\in S} x_i$ in its multilinear expansion. A low-degree bias means the learner minimizes this quantity, which orders hierarchical realizations $g\circ\Phi$ (a generic representation plus task-specific head) against flat realizations $h^*$ of each task. The machinery works through three lemmas: min-degree solutions of any task have degree at most $d$ (Lemma E.1), the averaged degree of $k$-degree functions composed with an invertible transform is minimized exactly when the transform is degree-1 (Lemma E.10), and conditioning on a representation lowers task degree only when the representation aligns with the latent structure. 'k-degree tasks'—tasks solvable by a function of degree at most $k$ on top of the true latents—are what make the task distribution break the uniform-case tie.

What would settle it

Train a multi-task MLP on $n$ tasks sampled from the degree-biased distribution of Theorem 4.8 with $p_1>0$, and linearly probe the learned representation for the true latents $z$: the theorem's mechanism predicts that the probe error tends to zero (up to sign and permutation) as $n$ grows, while single-task training ($n=1$) does not identify $z$. If the probe error stays bounded away from zero as $n\to\infty$ under exactly this setting, the identification claim as stated fails.

Watch

Extended reading notes

Core claim

The paper's claim is that the low-complexity bias of neural networks is enough to break latent-variable non-identifiability, provided the learner is trained on many tasks at once and the task distribution is tilted toward low-degree Boolean functions. Concretely, Theorem 4.8 states that as the number of tasks $n$ grows, the minimizer $\Phi^*$ of the average realization degree satisfies $\Phi^*_j(x)\in\{\pm z_{i_j}\}$ for each coordinate $j$: the learned representation coincides with the true latent variables up to signs and a permutation of coordinates, with $z=\psi^{-1}(x)$. This holds although the proxy tasks $h_i$ may be arbitrary nonlinear functions of $z$; the only structural requirement is that lower-degree tasks are sampled with higher probability and that degree-1 tasks have positive probability $p_1>0$. The authors complement this with a 'no free lunch' theorem (Theorem 4.4) showing that a uniform task distribution makes all representations equally cheap on average, and an architecture theorem (Theorem 4.12) showing that recovery persists exactly when the model's functional basis preserves degree ('basis compatibility').

Load-bearing premise

The result requires the task distribution to give positive probability to degree-1 tasks—proxy tasks that directly output an individual latent bit—and it assumes the learner exactly minimizes realization degree, which is not proven for networks trained by gradient descent.

Editorial extensions

If this is right

  • Single-task training never favors a world model: for every task, the flat min-degree solution is at least as cheap as any hierarchical one (Theorem 4.1), so representation learning requires the multi-task setting.
  • With enough tasks of positive conditional degree, hierarchical realizations beat flat ones (Theorem 4.3), giving a theoretical reason why pre-training objectives like next-token prediction or masked modeling can induce general representations.
  • A world-model representation is provably better out-of-distribution: for tasks whose latent-space version is a parity of degree $q$, flat min-degree solutions have test MSE above 1 while the world-model realization has test MSE 0 when training latents are restricted to a Hamming ball (Theorem 4.9), an abstract model of length generalization.
  • Architecture controls the outcome: only bases that preserve the degree of every parity function ('compatible' bases) inherit the recovery guarantee; for any $k$ there is an incompatible basis whose minimizer is $\Phi=T\circ\psi^{-1}$ with $\max_i \deg(T_i^{-1})\ge k$, so the wrong activation basis can hide the latents in high-degree features (Theorem 4.12).
  • The result is a Boolean, provable counterpart of the linear representation hypothesis: permutations and negations are exactly the degree-1 Boolean functions, the discrete analogue of linear features in real space.

Reading between the lines

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

  • If the mechanism is right, the practical recipe is to shape pretraining task mixtures toward low-degree tasks and to prefer objectives that raise conditional degree; reconstruction-style losses that are solvable by a low-degree shortcut (predicting pixel $x$ from $x$) should fail, while masked or predictive objectives should succeed—a comparative prediction the paper only gestures at.
  • The $p_1>0$ requirement yields a testable diagnostic: vary the fraction of degree-1 tasks in the mixture and measure latent-recovery error; the theory predicts a sharp transition in the number of tasks needed, and failure near $p_1=0$ would confirm that 'atomic' tasks are load-bearing rather than a technical convenience.
  • The exact-minimizer idealization suggests the cleanest place to look for the predicted effect is heavily regularized or feature-learning-dominated training that amplifies simplicity bias; networks that memorize via flat solutions should show no latent recovery.
  • By recovering latents only up to permutations and negations, the result meets the identifiability bar used in causal representation learning; adding interventional or auxiliary-variable signals could plausibly tighten the equivalence class to the exact latents, a direction the paper lists as future work but does not develop.
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 / 5 minor

Summary. The paper proposes a Boolean-function framework for studying when neural network representations recover latent data-generating variables, which it equates with learning world models. It defines a realization-degree complexity measure, proves that single-task learning prefers flat solutions (Theorem 4.1), that enough tasks with positive conditional degree can favor hierarchical solutions (Theorem 4.3), and that a uniform task distribution yields a no-free-lunch result (Theorem 4.4). The main result, Theorem 4.8, states that if tasks are drawn from a degree-weighted distribution over functions of the true latents with p1>0, then the exact minimizer of the average realization degree recovers the latents up to negation and permutation. The paper also claims out-of-distribution benefits (Theorem 4.9), analyzes architecture dependence via basis compatibility (Theorem 4.12), and reports experiments on polynomial extrapolation and physical-law prediction.

Significance. The central identifiability theorem (Theorem 4.8) is a genuinely interesting result: it gives a precise sense in which a low-degree bias in a multi-task curriculum can break generic non-identifiability of nonlinear latent-variable models. The proof relies on Lemma E.10, a clean averaging argument over degree-bounded Boolean function classes, and the equality analysis for k=1 is coherent. The paper also honestly states several limitations, including the p1>0 assumption and the counterexample in Example E.12. If the formal gaps are repaired, this could be a useful contribution to the theory of representation learning. However, the current version has serious issues in the OOD-benefit theorem and in the statement of Theorem 4.3, and the bridge from exact minimizers to neural network training is only empirical.

major comments (4)
  1. [Section 4.3 / Section E.7, Theorem 4.9] As stated, Theorem 4.9 is vacuous. Since h∘ψ is a parity function of degree q, the function g=χ_S satisfies g∘ψ^{-1}∈H(h), so max_g deg(g)≥q. Lemma E.13 gives deg(h*)≤k for every h*∈Hmin(h). By Definition 3.6, deg(h|ψ^{-1}) = deg(h*)−max_g deg(g) ≤ k−q. The assumptions state q>k and deg(h|ψ^{-1})≥q−r, but k−q<0<q−r, so no task can satisfy both. Additionally, the proof of part (ii) is invalid: the unique degree-r function from Lemma E.8 agrees with a degree-q parity only on the Hamming ball B_r, and since q>r it cannot equal the parity on all of Z; therefore the claimed zero OOD MSE for g*∘Φ* does not follow from uniqueness of low-degree interpolation.
  2. [Section 4.1 / Section E.3, Theorem 4.3] The statement of Theorem 4.3 quantifies over 'Φ* ∈ Hmin(Φ)', but Hmin(Φ) is never defined anywhere in the paper. The proof's bound deg(Φ*_j)≤d, imported from Lemma E.1, applies to min-degree solutions of tasks, not to arbitrary representations Φ*. Without a definition of the class over which Φ* minimizes, inequality (7) is not a well-formed theorem. The authors should either define Hmin(Φ) as the set of degree-minimizing representations subject to the feasibility constraints and prove the bound, or restate the result without this undefined object.
  3. [Section E.6, proof of Theorem 4.8] The proof asserts without justification that the feasibility constraints gi∘Φ∈H(hi) for all i imply the existence of a bijection T:Z→Z with Φ(x)=T(z) for every x=ψ(z). This is true in the n→∞ limit when p1>0, because the degree-1 tasks separate points of Z, but the argument is omitted and should be made explicit. Furthermore, Theorem 4.8 is a statement about exact minimizers of the realization-degree objective (10); the paper does not show that any neural network training procedure converges to such a minimizer. The abstract's claim that 'models with a low-degree bias provably recover' should therefore be qualified to refer to the idealized optimization problem, with the experimental section serving as empirical evidence.
  4. [Theorem 4.8 and Remark E.11, p1>0] The assumption p1>0 is substantive, not merely technical: it requires the task distribution to place positive probability on degree-1 tasks that directly output individual true latents. Example E.12 shows that without p1>0 the minimizing representation need not be a world model, so this assumption is not removable in the current framework. The paper acknowledges this in Remark E.11, but the abstract's phrase 'mild assumptions' overstates the scope. The abstract and introduction should explicitly mention that the result requires some direct linear (degree-1) observations of the latents, not merely nonlinear functions of them.
minor comments (5)
  1. [Section E.7, proof of Lemma E.13] The displayed formula 'Pr i=0 (d r)' should read '\sum_{i=0}^r \binom{d}{i}'; as printed it is not a valid expression.
  2. [Theorem 4.9] The notation 'deg(h)=q' in Theorem 4.9 is ambiguous; the subsequent conditions refer to deg(h∘ψ), so the statement should say 'deg(h∘ψ)=q'.
  3. [Remark 4.10] The phrase 'zero i.i.d. test error' is confusing because Theorem 4.9 defines the test distribution as uniform over Z, where the flat solution has error larger than 1. The intended meaning appears to be zero error on a held-out sample from the training distribution B_r; please reword.
  4. [Section 4.4, Eq. (11)] The definition of deg_U(f) in Eq. (11) seems to use the coefficients of f in the standard parity basis rather than the coefficients of f in the new basis {U(χ_S)}. Please clarify the definition, ideally with a concrete example, or correct the formula.
  5. [Figure 2c] The numeric labels in Figure 2c (0.16, 0.29, 0.07, 0.13) are not explained in the caption; please add a legend or a table so the reader knows which numbers correspond to which model and task.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 4.8's derivation is self-contained; the p1>0 requirement is a substantive but honestly acknowledged scope condition, not a circular encoding of the conclusion.

full rationale

The central claim (Theorem 4.8, Section 4.2; proof in Section E.6) is derived, not assumed: Lemma E.10 (Eqs. 51-60) proves a genuine inequality for sums of degrees under invertible transforms, and the equality condition deg(T_i)=1 for k=1 is proven, not posited. The task distribution being defined in terms of the true latents (Definition 4.6, F^d_k ∘ ψ^{-1}) is the standard form of an identifiability assumption, analogous to assuming sparsity in a known dictionary, not a self-definitional reduction: the theorem's contribution is showing that the degree-minimizer must be a degree-1 transform of z, which Example E.12 shows is false without p1>0. The paper explicitly flags the p1>0 limitation (Remark E.11 and the bullet after Theorem 4.8: 'we require a non-zero probability of explicitly sampling degree-1 tasks'), and Example E.12 with T=(z1, z1 z2, z1 z3) demonstrates the condition is essential; this is honest scope disclosure rather than hidden circularity. No parameters are fitted and then renamed as predictions: the optimization (10) is over exact realization degree, and the asymptotic n→∞ analysis is analytic. Self-citations (Zhang et al. 2024; Zhao et al. 2024) occur only in the related-work survey (Appendix A) and are not load-bearing for any theorem. Lemma E.8, the one imported uniqueness result, comes from external prior work (Abbe et al. 2023, Theorem 5.1) and is cited as an external mathematical fact, not from the present authors. The abstract's phrase 'mild assumptions' arguably overstates the scope, since p1>0 requires some proxy tasks to output individual latents directly, but that is a scope/correctness concern, not circularity. The empirical sections are proof-of-concept demonstrations, not evidence used in the proofs. Overall, the derivation chain is self-contained and no step reduces by construction to its own inputs.

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

The central theorems rest on the Boolean encoding, the invertible data-generation process, the assumed low-degree minimization objective, and the degree-biased task distribution with p1 > 0. These are stated assumptions rather than derived facts; they are the price of the identifiability result.

free parameters (2)
  • task degree distribution p_1,...,p_d = not fitted; p_i in (0,1), sum = 1
    Theorem 4.8's identifiability result depends on this distribution; the critical constraint p1 > 0 is required for recovery up to negations and permutations, and Example E.12 shows failure when p1 = 0.
  • Hamming ball radius r = r < d, no fitted value
    Theorem 4.9's OOD benefit is stated for training latents in B_r; the degree bound k and the conditions deg(h|ψ^{-1}) ≥ q-r depend on r. It is a problem parameter, not fitted.
assumptions (5)
  • domain assumption All variables can be modeled as Boolean without loss of generality because computers use finite precision.
    Invoked in Section 3 to justify replacing continuous latents and data with bit strings; it underlies all Fourier-Walsh degree analysis.
  • domain assumption The data generation process is x = ψ(z) with invertible ψ and supp(p) = Z.
    Definitions 2.1 and Section 3; invertibility gives |X| = |Z| and supports the bijective-transform reduction in Theorem 4.8.
  • domain assumption The learner exhibits a low-degree bias, i.e., minimizes realization degree ddeg over realizations.
    Definition 3.4 and Section 4 state this; the theorems characterize the minimizer of Eq. (10), not the output of an actual training algorithm.
  • ad hoc to paper Proxy tasks are sampled from a degree-weighted distribution over F_d ∘ ψ^{-1} with p_i in (0,1); in particular p1 > 0.
    Definition 4.6 and Theorem 4.8; this distribution encodes the latent coordinates as the simple coordinates. Example E.12 shows the conclusion fails for a natural nonlinear T when p1 = 0.
  • standard math Lemma E.8 from Abbe et al. (2023): on Hamming ball B_r there is a unique degree ≤ r Boolean function matching any target.
    Used in Theorem 4.9(ii); cited, not proved in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Do Neural Networks Learn World Models?." pith.science (2026). https://pith.science/paper/GAXBHNQM

@misc{pith2026250209297,
  author       = {Pith},
  title        = {Pith review of: When Do Neural Networks Learn World Models?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GAXBHNQM}},
  note         = {Machine review of arXiv:2502.09297}
}
read the original abstract

Humans develop world models that capture the underlying generation process of data. Whether neural networks can learn similar world models remains an open problem. In this work, we present the first theoretical results for this problem, showing that in a multi-task setting, models with a low-degree bias provably recover latent data-generating variables under mild assumptions--even if proxy tasks involve complex, non-linear functions of the latents. However, such recovery is sensitive to model architecture. Our analysis leverages Boolean models of task solutions via the Fourier-Walsh transform and introduces new techniques for analyzing invertible Boolean transforms, which may be of independent interest. We illustrate the algorithmic implications of our results and connect them to related research areas, including self-supervised learning, out-of-distribution generalization, and the linear representation hypothesis in large language models.

Figures

Figures reproduced from arXiv: 2502.09297 by the authors.

Figure 1
Figure 1. A graphical summary of our framework and main results. laws (Kang et al., 2025). We show that architectures inspired by our analysis outperform conventional archi￾tectures such as ReLU MLPs and transformers (Vaswani et al., 2017) in these tasks. 2. Formulation of Learning World Models How to define the world model and the problem of learning world models remains debatable to date. Yet, the term “world models” has be… view at source ↗
Figure 2
Figure 2. Empirical results. (a) An example of extrapolating a degree-3 polynomial. Shaded region indicates the training region. (b) Violin plots of the test mean square error (MSE) of ReLU MLPs and our models in extrapolating degree-2 (left) and degree-3 (right) polynomials. (c) Results for learning physical laws. Each column indicates the task and out-of-distribution test MSE averaged over 5 runs. analyzing similar concepts… view at source ↗
Figure 3
Figure 3. Low-degree bias of neural networks: Training latents are uniformly sampled from the Hamming ball Br = {z ∈ {±1} d | #−1(z) ≤ r}; the out-of-distribution (OOD) test latents are uniformly sampled from {±1} d (d = 10). For small r, low-degree realizations of tasks (target functions) exist in the training distribution. A model that learns these low-degree realizations would have low in-distribution (ID) error yet high O… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Impact of multi-task training on identifying data-generating variables: The identification error of data-generating variables decreases as the task number n increases, measured by the MSE of the linear probe trained to fit true latent variables z given the learned repr…
Figure 5
Figure 5. Figure 5: Identifying data-generating variables benefits out-of-distribution (OOD) generalization: For the same tasks considered in [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Two visualized examples of the parabolic motion. F.1. Polynomial Extrapolation Dataset. We consider fitting and extrapolating degree-n polynomials with the form Pn(x) = Pn i=0 aix n. Given an input x ∈ R, the label is given by y = Pn(x). In our experiments, we consider…
Figure 7
Figure 7. Figure 7: Two visualized examples of the collision motion. F.2. Learning Physical Laws Dataset. Inspired by Kang et al. (2025), we create training and test sequences representing ball-shaped object movements that adhere to two basic physical laws: (i) single-object parabolic mot…
Figure 8
Figure 8. Figure 8: Violin plots of the test MSE of the ReLU MLP and our model in extrapolating degree-1 polynomials. We train all models using the MSE loss with the AdamW optimizer. Training hyperparameters are as follows: • Initial learning rate is randomly sampled from [1e − 6, 1e − 3]…
Figure 9
Figure 9. Figure 9: Selected examples for degree-1 polynomial extrapolation. Shaded regions indicate training regions. Ground Truth ReLU MLP Ours (a) Ground Truth ReLU MLP Ours (b) Ground Truth ReLU MLP Ours (c) [PITH_FULL_IMAGE:figures/full_fig_p031_9.png]
Figure 10
Figure 10. Figure 10: Selected examples for degree-2 polynomial extrapolation. Shaded regions indicate training regions. Ground Truth ReLU MLP Ours (a) Ground Truth ReLU MLP Ours (b) Ground Truth ReLU MLP Ours (c) [PITH_FULL_IMAGE:figures/full_fig_p031_10.png]
Figure 11
Figure 11. Figure 11: Selected examples for degree-3 polynomial extrapolation. Shaded regions indicate training regions. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_11.png]
Figure 12
Figure 12. Figure 12: Visualization results in a test collision motion example. All three rows select the same frames with uniform spacing. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

95 extracted references · 59 canonical work pages

  1. [1]

    write newline

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

  2. [2]

    Generalization on the unseen, logic reasoning and degree curriculum

    Abbe, E., Bengio, S., Lotfi, A., and Rizk, K. Generalization on the unseen, logic reasoning and degree curriculum. In International Conference on Machine Learning , 2023

    2023

  3. [3]

    Interventional causal representation learning

    Ahuja, K., Mahajan, D., Wang, Y., and Bengio, Y. Interventional causal representation learning. In International Conference on Machine Learning , 2023

    2023

  4. [4]

    and Li, Y

    Allen-Zhu, Z. and Li, Y. Towards understanding ensemble, knowledge distillation and self-distillation in deep learning. In International Conference on Learning Representations , 2023

    2023

  5. [5]

    V., Pillaud-Vivien, L., and Flammarion, N

    Andriushchenko, M., Varre, A. V., Pillaud-Vivien, L., and Flammarion, N. SGD with large step sizes learns sparse features. In International Conference on Machine Learning , pp.\ 903--925, 2023

    2023

  6. [6]

    Exploring length generalization in large language models

    Anil, C., Wu, Y., Andreassen, A., Lewkowycz, A., Misra, V., Ramasesh, V., Slone, A., Gur-Ari, G., Dyer, E., and Neyshabur, B. Exploring length generalization in large language models. In Advances in Neural Information Processing Systems , volume 35, pp.\ 38546--38556, 2022

    2022

  7. [7]

    Invariant risk minimization

    Arjovsky, M., Bottou, L., Gulrajani, I., and Lopez-Paz, D. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019

    arXiv 1907

  8. [8]

    A theoretical analysis of contrastive unsupervised representation learning

    Arora, S., Khandeparkar, H., Khodak, M., Plevrakis, O., and Saunshi, N. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning , pp.\ 5628--5637, 2019

    2019

Show all 95 references
  1. [9]

    Bartlett, P. L. and Mendelson, S. Rademacher and Gaussian complexities: Risk bounds and structural results. Journal of Machine Learning Research, 3: 0 463--482, 2002

    2002

  2. [10]

    L., Foster, D

    Bartlett, P. L., Foster, D. J., and Telgarsky, M. Spectrally-normalized margin bounds for neural networks. In Advances in Neural Information Processing Systems , 2017

    2017

  3. [11]

    L., Long, P

    Bartlett, P. L., Long, P. M., Lugosi, G., and Tsigler, A. Benign overfitting in linear regression. Proceedings of the National Academy of Sciences, 117 0 (48): 0 30063--30070, 2020

    2020

  4. [12]

    M., Gebru, T., McMillan-Major, A., and Shmitchell, S

    Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness , Accountability , and Transparency , pp.\ 610--623, 2021

    2021

  5. [13]

    Representation learning: A review and new perspectives

    Bengio, Y., Courville, A., and Vincent, P. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35 0 (8): 0 1798--1828, 2013

    2013

  6. [14]

    fail to learn ``b is a

    Berglund, L., Tong, M., Kaufmann, M., Balesni, M., Stickland, A. C., Korbak, T., and Evans, O. The reversal curse: LLMs trained on ``a is b" fail to learn ``b is a". In International Conference on Learning Representations , 2024

    2024

  7. [15]

    Simplicity bias in transformers and their ability to learn sparse boolean functions

    Bhattamishra, S., Patel, A., Kanade, V., and Blunsom, P. Simplicity bias in transformers and their ability to learn sparse boolean functions. arXiv preprint arXiv:2211.12316, 2023

    2023 arXiv

  8. [16]

    Towards monosemanticity: Decomposing language models with dictionary learning

    Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., et al. Towards monosemanticity: Decomposing language models with dictionary learning. Transformer Circuits Thread, 2, 2023

    2023

  9. [17]

    Video generation models as world simulators, 2024

    Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., et al. Video generation models as world simulators, 2024

    2024

  10. [18]

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., L...

    2020

  11. [19]

    and Sudijono, T

    Chatterjee, S. and Sudijono, T. Neural networks generalize on low complexity data. arXiv preprint arXiv:2409.12446, 2024

    2024

  12. [20]

    Implicit bias of gradient descent for wide two-layer neural networks trained with the logistic loss

    Chizat, L., Chizat, L., and Fr, U.-P.-S. Implicit bias of gradient descent for wide two-layer neural networks trained with the logistic loss. In Conference on Learning Theory , 2020

    2020

  13. [21]

    Craik, K. J. W. The nature of explanation, volume 445. CUP Archive, 1967

    1967

  14. [22]

    Q., and Louis, A

    Dingle, K., Camargo, C. Q., and Louis, A. A. Input–output maps are strongly biased towards simple outputs. Nature Communications, 9 0 (1), 2018

    2018

  15. [23]

    An introduction to latent variable models

    Everett, B. An introduction to latent variable models. 2013

    2013

  16. [24]

    J., Nagai, Y., Taniguchi, T., Gomi, H., and Tenenbaum, J

    Friston, K., Moran, R. J., Nagai, Y., Taniguchi, T., Gomi, H., and Tenenbaum, J. World model learning and inference. Neural Networks, 144: 0 573--590, 2021

    2021

  17. [25]

    On the approximate realization of continuous mappings by neural networks

    Funahashi, K.-I. On the approximate realization of continuous mappings by neural networks. Neural Networks, 2 0 (3): 0 183--192, 1989

    1989

  18. [26]

    A., and Brendel, W

    Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A., and Brendel, W. ImageNet -trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations , 2019

    2019

  19. [27]

    W., and Keutzer, K

    Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M. W., and Keutzer, K. A survey of quantization methods for efficient neural network inference. arXiv preprint arXiv:2103.13630, 2021

    2021 arXiv

  20. [28]

    Goldblum, M., Finzi, M., Rowan, K., and Wilson, A. G. Position: The no free lunch theorem, Kolmogorov complexity, and the role of inductive biases in machine learning. In International Conference on Machine Learning , 2024

    2024

  21. [29]

    and Bengio, Y

    Goyal, A. and Bengio, Y. Inductive biases for deep learning of higher-level cognition. arXiv preprint arXiv:2011.15091, 2020

    2011 arXiv

  22. [30]

    Characterizing implicit bias in terms of optimization geometry

    Gunasekar, S., Lee, J., Soudry, D., and Srebro, N. Characterizing implicit bias in terms of optimization geometry. In International Conference on Machine Learning , pp.\ 1827--1836, 2018 a

    2018

  23. [31]

    D., Soudry, D., and Srebro, N

    Gunasekar, S., Lee, J. D., Soudry, D., and Srebro, N. Implicit bias of gradient descent on linear convolutional networks. In Advances in Neural Information Processing Systems , pp.\ 9482--9491, 2018 b

    2018

  24. [32]

    and Tegmark, M

    Gurnee, W. and Tegmark, M. Language models represent space and time. In International Conference on Learning Representations , 2024

    2024

  25. [33]

    and Schmidhuber, J

    Ha, D. and Schmidhuber, J. World models. arXiv preprint arXiv:1803.10122, 2018

    2018 arXiv

  26. [34]

    Masked autoencoders are scalable vision learners

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., and Girshick, R. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition , pp.\ 16000--16009, 2022

    2022

  27. [35]

    An overview of catastrophic AI risks

    Hendrycks, D., Mazeika, M., and Woodside, T. An overview of catastrophic AI risks. arXiv preprint arXiv:2306.12001, 2023

    2023 arXiv

  28. [36]

    Multilayer feedforward networks are universal approximators

    Hornik, K., Stinchcombe, M., and White, H. Multilayer feedforward networks are universal approximators. Neural Networks , 2 0 (5): 0 359--366, 1989

    1989

  29. [37]

    The low-rank simplicity bias in deep networks

    Huh, M., Mobahi, H., Zhang, R., Cheung, B., Agrawal, P., and Isola, P. The low-rank simplicity bias in deep networks. Transactions on Machine Learning Research, 2023

    2023

  30. [38]

    The Platonic representation hypothesis

    Huh, M., Cheung, B., Wang, T., and Isola, P. The Platonic representation hypothesis. In International Conference on Machine Learning , 2024

    2024

  31. [39]

    and Pajunen, P

    Hyvärinen, A. and Pajunen, P. Nonlinear independent component analysis: Existence and uniqueness results. Neural Networks, 12 0 (3): 0 429--439, 1999

    1999

  32. [40]

    Hyvärinen, A., Sasaki, H., and Turner, R. E. Nonlinear ICA using auxiliary variables and generalized contrastive learning. In International Conference on Artificial Intelligence and Statistics , 2019

    2019

  33. [41]

    Neural tangent kernel: Convergence and generalization in neural networks

    Jacot, A., Gabriel, F., and Hongler, C. Neural tangent kernel: Convergence and generalization in neural networks. In Advances in Neural Information Processing Systems , 2018

    2018

  34. [42]

    Low -resource

    Jiang, Z., Yang, M., Tsirlin, M., Tang, R., Dai, Y., and Lin, J. “ Low -resource” text classification: A parameter-free classification method with compressors. In Findings of the Association for Computational Linguistics : ACL 2023 , pp.\ 6810--6828, Toronto, Canada, 2023

    2023

  35. [43]

    and Rinard, M

    Jin, C. and Rinard, M. Emergent representations of program semantics in language models trained on programs. In International Conference on Machine Learning , 2024

    2024

  36. [44]

    SGD on neural networks learns functions of increasing complexity

    Kalimeris, D., Kaplun, G., Nakkiran, P., Edelman, B., Yang, T., Barak, B., and Zhang, H. SGD on neural networks learns functions of increasing complexity. In Advances in Neural Information Processing Systems , pp.\ 3491--3501, 2019

    2019

  37. [45]

    How far is video generation from world model: A physical law perspective

    Kang, B., Yue, Y., Lu, R., Lin, Z., Zhao, Y., Wang, K., Huang, G., and Feng, J. How far is video generation from world model: A physical law perspective. In International Conference on Machine Learning , 2025

    2025

  38. [46]

    P., Monti, R

    Khemakhem, I., Kingma, D. P., Monti, R. P., and Hyvärinen, A. Variational autoencoders and nonlinear ICA : A unifying framework. In International Conference on Artificial Intelligence and Statistics , 2020

    2020

  39. [47]

    A path towards autonomous machine intelligence version 0.9

    LeCun, Y. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review, 62 0 (1): 0 1--62, 2022

    2022

  40. [48]

    D., Lei, Q., Saunshi, N., and Zhuo, J

    Lee, J. D., Lei, Q., Saunshi, N., and Zhuo, J. Predicting what you already know helps: Provable self-supervised learning. In Advances in Neural Information Processing Systems , volume 34, pp.\ 309--323, 2021

    2021

  41. [49]

    Z., Nye, M., and Andreas, J

    Li, B. Z., Nye, M., and Andreas, J. Implicit representations of meaning in neural language models. arXiv preprint arXiv:2106.00737, 2021

    2021 arXiv

  42. [50]

    K., and Bau, D

    Li, K., Hopkins, A. K., and Bau, D. Emergent world representations: Exploring a sequence model trained on a synthetic task. In International Conference on Learning Representations , 2023

    2023

  43. [51]

    An introduction to Kolmogorov complexity and its applications, volume 3

    Li, M., Vit \'a nyi, P., et al. An introduction to Kolmogorov complexity and its applications, volume 3. 2008

    2008

  44. [52]

    Self-supervised learning: Generative or contrastive

    Liu, X., Zhang, F., Hou, Z., Mian, L., Wang, Z., Zhang, J., and Tang, J. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowleadge and Data Engineering , 35 0 (1): 0 857--876, 2021

    2021

  45. [53]

    Grokking as compression: A nonlinear complexity perspective

    Liu, Z., Zhong, Z., and Tegmark, M. Grokking as compression: A nonlinear complexity perspective. arXiv preprint arXiv:2310.05918, 2023

    2023 arXiv

  46. [54]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. In International Conference on Learning Representations , 2019

    2019

  47. [55]

    Lotfi, S., Finzi, M., Kapoor, S., Potapczynski, A., Goldblum, M., and Wilson, A. G. PAC - Bayes compression bounds so tight that they can explain generalization. In Advances in Neural Information Processing Systems , 2022

    2022

  48. [56]

    Gradient descent on two-layer nets: Margin maximization and simplicity bias

    Lyu, K., Wang, R., Li, Z., and Arora, S. Gradient descent on two-layer nets: Margin maximization and simplicity bias. In Advances in Neural Information Processing Systems , pp.\ 12978--12991, 2021

    2021

  49. [57]

    and Tegmark, M

    Marks, S. and Tegmark, M. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824, 2023

    2023 arXiv

  50. [58]

    Linguistic regularities in continuous space word representations

    Mikolov, T., Yih, W.-t., and Zweig, G. Linguistic regularities in continuous space word representations. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics : Human Language Technologies , pp.\ 746--751, 2013

    2013

  51. [59]

    GSM -symbolic: Understanding the limitations of mathematical reasoning in large language models

    Mirzadeh, I., Alizadeh, K., Shahrokhi, H., Tuzel, O., Bengio, S., and Farajtabar, M. GSM -symbolic: Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229, 2024

    2024 arXiv

  52. [60]

    Ai’s challenge of understanding the world

    Mitchell, M. Ai’s challenge of understanding the world. Science, 382 0 (6671), 2023

    2023

  53. [61]

    Do generative video models learn physical principles from watching videos? arXiv preprint arXiv:2501.09038, 2025

    Motamed, S., Culp, L., Swersky, K., Jaini, P., and Geirhos, R. Do generative video models learn physical principles from watching videos? arXiv preprint arXiv:2501.09038, 2025

    2025 arXiv

  54. [62]

    A., Bondarenko, Y., Baalen, M

    Nagel, M., Fournarakis, M., Amjad, R. A., Bondarenko, Y., Baalen, M. v., and Blankevoort, T. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021

    2021 arXiv

  55. [63]

    Emergent linear representations in world models of self-supervised sequence models

    Nanda, N., Lee, A., and Wattenberg, M. Emergent linear representations in world models of self-supervised sequence models. arXiv preprint arXiv:2309.00941, 2023

    2023 arXiv

  56. [64]

    Analysis of boolean functions

    O'Donnell, R. Analysis of boolean functions. Cambridge University Press, 2014

    2014

  57. [65]

    PyTorch : An imperative style, high-performance deep learning library

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., and Chintala, S. PyTorch : An imper...

    2019

  58. [66]

    Train short, test long: Attention with linear biases enables input length extrapolation

    Press, O., Smith, N., and Lewis, M. Train short, test long: Attention with linear biases enables input length extrapolation. In International Conference on Learning Representations , 2022

    2022

  59. [67]

    V., Louis, A

    Pérez, G. V., Louis, A. A., and Camargo, C. Q. Deep learning generalizes because the parameter-function map is biased towards simple functions. In International Conference on Learning Representations , 2019

    2019

  60. [68]

    Language models are unsupervised multitask learners

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. OpenAI Blog , 1 0 (8): 0 9, 2019

    2019

  61. [69]

    Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems , 34: 0 12116--12128, 2021

    Raghu, M., Unterthiner, T., Kornblith, S., Zhang, C., and Dosovitskiy, A. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems , 34: 0 12116--12128, 2021

    2021

  62. [70]

    Position: Understanding LLMs requires more than statistical generalization

    Reizinger, P., Ujváry, S., Mészáros, A., Kerekes, A., Brendel, W., and Huszár, F. Position: Understanding LLMs requires more than statistical generalization. In International Conference on Machine Learning , 2024

    2024

  63. [71]

    and Everitt, T

    Richens, J. and Everitt, T. Robust agents learn causal world models. In ICLR, 2024

    2024

  64. [72]

    R., Kalchbrenner, N., Goyal, A., and Bengio, Y

    Schölkopf, B., Locatello, F., Bauer, S., Ke, N. R., Kalchbrenner, N., Goyal, A., and Bengio, Y. Toward causal representation learning. Proceedings of the IEEE, 109 0 (5): 0 612--634, 2021

    2021

  65. [73]

    S., Gunasekar, S., and Srebro, N

    Soudry, D., Hoffer, E., Nacson, M. S., Gunasekar, S., and Srebro, N. The implicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 19 0 (1): 0 2822--2878, 2018

    2018

  66. [74]

    Neural redshift: Random networks are not random functions

    Teney, D., Nicolicioiu, A., Hartmann, V., and Abbasnejad, E. Neural redshift: Random networks are not random functions. arXiv preprint arXiv:2403.02241, 2024

    2024 arXiv

  67. [75]

    Contrastive learning, multi-view redundancy, and linear models

    Tosh, C., Krishnamurthy, A., and Hsu, D. Contrastive learning, multi-view redundancy, and linear models. In Algorithmic Learning Theory , pp.\ 1179--1206, 2021

    2021

  68. [76]

    The nature of statistical learning theory

    Vapnik, V. The nature of statistical learning theory. 1999

    1999

  69. [77]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need. In Advances in Neural Information Processing Systems , pp.\ 5998--6008, 2017

    2017

  70. [78]

    Self-supervised learning with data augmentations provably isolates content from style

    Von Kügelgen, J., Sharma, Y., Gresele, L., Brendel, W., Schölkopf, B., Besserve, M., and Locatello, F. Self-supervised learning with data augmentations provably isolates content from style. In Advances in Neural Information Processing Systems , volume 34, pp.\ 16451--16467, 2021

    2021

  71. [79]

    Nonparametric identifiability of causal representations from unknown interventions

    von Kügelgen, J., Besserve, M., Wendong, L., Gresele, L., Kekić, A., Bareinboim, E., Blei, D., and Schölkopf, B. Nonparametric identifiability of causal representations from unknown interventions. In Advances in Neural Information Processing Systems , volume 36, 2023

    2023

  72. [80]

    M., and Ma, T

    Wei, C., Xie, S. M., and Ma, T. Why do pretrained language models help in downstream tasks? An analysis of head and prompt tuning. In Advances in Neural Information Processing Systems , volume 34, pp.\ 16158--16170, 2021

    2021

  73. [81]

    and Watson, J

    Whitley, D. and Watson, J. P. Complexity theory and the no free lunch theorem. Search Methodologies : Introductory Tutorials in Optimization and Decision Support Techniques , pp.\ 317--339, 2005

    2005

  74. [82]

    Wolpert, D. H. The lack of a priori distinctions between learning algorithms. Neural computation, 8 0 (7): 0 1341--1390, 1996

    1996

  75. [83]

    K., Goodman, N

    Wong, L., Grand, G., Lew, A. K., Goodman, N. D., Mansinghka, V. K., Andreas, J., and Tenenbaum, J. B. From word models to world models: Translating from natural language to the probabilistic language of thought. arXiv preprint arXiv:2306.12672, 2023

    2023 arXiv

  76. [84]

    Reasoning or reciting? Exploring the capabilities and limitations of language models through counterfactual tasks

    Wu, Z., Qiu, L., Ross, A., Akyürek, E., Chen, B., Wang, B., Kim, N., Andreas, J., and Kim, Y. Reasoning or reciting? Exploring the capabilities and limitations of language models through counterfactual tasks. arXiv preprint arXiv:2307.02477, 2023

    2023 arXiv

  77. [85]

    Making large language models into world models with precondition and effect knowledge

    Xie, K., Yang, I., Gunerli, J., and Riedl, M. Making large language models into world models with precondition and effect knowledge. arXiv preprint arXiv:2409.12278, 2024

    2024 arXiv

  78. [86]

    S., Kawarabayashi, K.-i., and Jegelka, S

    Xu, K., Li, J., Zhang, M., Du, S. S., Kawarabayashi, K.-i., and Jegelka, S. What can neural networks reason about? In International Conference on Learning Representations , 2020

    2020

  79. [87]

    S., Kawarabayashi, K.-i., and Jegelka, S

    Xu, K., Li, J., Zhang, M., Du, S. S., Kawarabayashi, K.-i., and Jegelka, S. How neural networks extrapolate: From feedforward to graph neural networks. In International Conference on Learning Representations , 2021

    2021

  80. [88]

    J., Zhang, Y., Luo, T., Xiao, Y., and Ma, Z

    Xu, Z.-Q. J., Zhang, Y., Luo, T., Xiao, Y., and Ma, Z. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019

    1901 arXiv

  81. [89]

    and Paul, L

    Yildirim, I. and Paul, L. From task structures to world models: What do LLMs know? Trends in Cognitive Sciences, 2024

    2024

  82. [90]

    Understanding deep learning requires rethinking generalization

    Zhang, C., Bengio, S., Hardt, M., Recht, B., and Vinyals, O. Understanding deep learning requires rethinking generalization. In International Conference on Learning Representations , 2017

    2017

  83. [91]

    Feature contamination: Neural networks learn uncorrelated features and fail to generalize

    Zhang, T., Zhao, C., Chen, G., Jiang, Y., and Chen, F. Feature contamination: Neural networks learn uncorrelated features and fail to generalize. In International Conference on Machine Learning , 2024

    2024

  84. [92]

    M ^3 PL : Identifying and exploiting view bias of prompt learning

    Zhao, C., Zhang, T., Chen, G., Jiang, Y., and Chen, F. M ^3 PL : Identifying and exploiting view bias of prompt learning. Transactions on Machine Learning Research, 2024

    2024

  85. [93]

    P., and Orbanz, P

    Zhou, W., Veitch, V., Austern, M., Adams, R. P., and Orbanz, P. Non-vacuous generalization bounds at the ImageNet scale: A PAC - Bayesian compression approach. In International Conference on Learning Representations , 2019

    2019

  86. [94]

    S., Sharma, Y., Schneider, S., Bethge, M., and Brendel, W

    Zimmermann, R. S., Sharma, Y., Schneider, S., Bethge, M., and Brendel, W. Contrastive learning inverts the data generating process. In International Conference on Machine Learning , pp.\ 12979--12990, 2021

    2021

  87. [95]

    Neural networks fail to learn periodic functions and how to fix it

    Ziyin, L., Hartwig, T., and Ueda, M. Neural networks fail to learn periodic functions and how to fix it. In Advances in Neural Information Processing Systems , 2020

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.