Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Learning Distributions over Permutations and Rankings with Factorized Representations

T0 review · 4 major / 4 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 Permutation codes turn ranking into ordinary token prediction, the paper claims.

desk verdict The representation theory is clean and the jigsaw results are genuinely strong, but the paper's headline claim about learning non-trivial distributions at low compute is not actually measured by the benchmarks as presented. read the letter →

arxiv 2505.24664 v1 pith:JRF3I4EW submitted 2025-05-30 cs.LG

Daniel Severo , Brian Karrer , Niklas Nolte This is my paper
classification cs.LG
keywords permutationdistributionsLehmercodesFisher-YatesshuffleinsertionvectorsRepeatedModelmaskedlanguagemodelingMallowsranking
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 aims to show that learning distributions over permutations can be reduced to token prediction over three factorized codes: Lehmer codes, Fisher-Yates draws, and insertion vectors, each of which maps permutations bijectively to integer sequences. Because every in-domain code sequence decodes to a valid permutation, a model trained with ordinary masked or autoregressive cross-entropy cannot output invalid permutations and can in principle represent any distribution on the symmetric group. In the fully factorized single-forward-pass setting, the smallest such family still contains classical models like Mallows and the Repeated Insertion Model, whereas inline permutation notation collapses to a delta distribution. On the CIFAR-10 jigsaw puzzle benchmark the reported accuracy is higher than previous diffusion and convex-relaxation baselines, and the paper introduces two benchmarks, cyclic permutations and MovieLens re-ranking, where non-trivial distributions are learned and simple baselines fail to produce valid permutations.

What carries the argument

The machinery is the family of factorized representations: the right-Lehmer code $L(X)_i$ counts smaller elements to the right of $X_i$; the Fisher-Yates draws record the swap offsets of the shuffle that produces $X$; and the insertion vector $V(X)_k$ records the slot, relative to a reference ranking, into which the $k$-th reference item is inserted. The load-bearing property is that each code is a bijection with the symmetric group, so the support constraint that all outputs must be valid permutations is enforced purely by the token domains, not by a learned constraint. The paper also derives Theorem 4.3, $V(X)_k = k - L(X^{-1})_k$, which reduces insertion-vector decoding to a Lehmer-code decode of the inverse permutation and makes batch decoding practical.

What would settle it

Measure the model's full joint distribution at one forward pass on the cyclic-permutation benchmark: if the fully-factorized Fisher-Yates model does not concentrate on cyclic permutations and does not approach uniform mass over the held-out cycles as training grows, the claim that factorized representations learn non-trivial distributions at one NFE is falsified. A direct check is the KL divergence from the fitted fully-factorized code distribution to a uniform distribution over the $(n-1)$ cyclic permutations of a fixed length $n$; a value that fails to decrease with data would contradict the paper's central claim.

Watch

Extended reading notes

Core claim

The central claim is that three classic permutation encodings are not just convenient algorithms but interchangeable coordinate systems for the symmetric group that make unconstrained neural density estimation over permutations possible. The paper argues that the right-Lehmer code, the Fisher-Yates draw sequence, and the insertion vector each give a bijection between permutations and integer sequences with position-dependent ranges, so any probability mass function over permutations corresponds to some distribution over these codes and vice versa. It further shows that in the autoregressive limit these factorized models subsume Mallows' weighted model and the Repeated Insertion Model, and at one function evaluation they still describe non-trivial families instead of collapsing. A new identity, Theorem 4.3, states that $V(X)_k = k - L(X^{-1})_k$, connecting insertion vectors to left-Lehmer codes of the inverse permutation and enabling batched decoding of insertion vectors through Lehmer machinery.

Load-bearing premise

The load-bearing premise is that cross-entropy-trained conditional distributions over code tokens, including those sampled in parallel at low numbers of forward passes, approximate the target distribution on held-out data; the paper demonstrates this empirically on three benchmarks but provides no theoretical guarantee for the low-NFE regime.

Editorial extensions

If this is right

  • At one forward pass, factorized representations can still represent non-trivial distributions such as Mallows' model and the Repeated Insertion Model, whereas any inline model restricted to one pass can only output a delta distribution.
  • Every sample from the model is a valid permutation by construction, at any number of forward passes, eliminating the need for constrained decoding or projection onto the permutation polytope.
  • On jigsaw puzzles the method reports higher accuracy than diffusion and Gumbel-Sinkhorn baselines, with masked models solving the puzzle in a single function evaluation.
  • Because insertion vectors condition naturally on sub-rankings, the same trained model can predict the remaining ranking given any observed prefix, which the MovieLens experiment exploits.
  • Theorem 4.3 makes insertion-vector sampling fast enough for batched inference, removing a computational obstacle to using this representation.

Reading between the lines

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

  • A testable extension is to plug these codes into other generative training objectives, such as discrete diffusion or flow matching, rather than only masked and autoregressive losses; the bijection property should carry over and may give smoother low-temperature sampling.
  • The identity in Theorem 4.3 suggests that the left-Lehmer code of the inverse permutation is exactly the insertion vector, so any algorithm or metric developed for Lehmer codes transfers immediately to insertion vectors; this could simplify future analyses of Repeated Insertion Model-type families.
  • If the low-NFE family characterization from Section 6 is filled in, it may show that the one-NFE Fisher-Yates model for cyclic permutations is exactly the uniform distribution over cycles, since the positivity condition $F Y_i > 0$ is necessary and sufficient for cyclicity.
  • The MovieLens setting hints that the same approach could be applied to ranking with partial preference data in recommender systems, where the item set is not fixed in advance; the paper notes this as future work.
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 / 4 minor

Summary. This paper proposes to learn distributions over permutations by training language models on factorized representations—right-Lehmer codes, Fisher-Yates draws, and insertion vectors—each of which is in bijection with the symmetric group. The authors argue that modelling these representations rather than the inline notation avoids the capacity collapse of fully factorized inline models, gives automatic validity guarantees, and subsumes Mallows and RIM families in appropriate conditional limits. They also prove a relationship between insertion vectors and left-Lehmer codes of inverse permutations (Theorem 4.3) to support batched decoding. Empirically, the paper reports strong results on CIFAR-10 jigsaw puzzles and introduces two new benchmarks: uniform cyclic permutations of length 10 and MovieLens re-ranking at sizes up to n=50.

Significance. The theoretical framework is attractive: the representations are simple, the bijections are standard but well packaged for deep learning, and the low-NFE compute/expressivity trade-off is practically useful. The paper includes working encode/decode code for all three representations, constructive subsumption arguments for Mallows and RIM, and a new decoding identity that, once indexing is clarified, enables batched insertion-vector decoding. The jigsaw results are strong and the new benchmarks are welcome in principle. However, the experiments designed to demonstrate learning of non-trivial distributions do not measure distributional fidelity, and two theoretical statements contain errors that need correction before the central claims are fully supported.

major comments (4)
  1. [Section 4.1, Eq. (4)] Equation (4) does not bound the entropy of inline models as stated. For n=4 with partition S1={1,2}, S2={3,4}, define X1 uniform on {1,2}, X2 uniform on {3,4}, and X3,X4 equal to the two remaining values in increasing order. This distribution satisfies the factorization in Eq. (2), assigns probability only to valid permutations, and has entropy 2 bits; the right-hand side of Eq. (4) is log2(3)+log2(1)=1.585 bits. The 1-NFE delta conclusion is correct, but the general capacity-loss argument for intermediate NFEs needs a corrected bound.
  2. [Section B.2, Theorem 4.3] In the proof of Theorem 4.3, Eq. (7) is false with the paper's stated 1-based index sets. For X=[3,5,4,1,2] and k=5, k−L(X)_5=5−3=2, while |{j<5: X_j<X_5}|=1. The identity V(X)_k=k−L(X^{-1})_k is correct when k and sequence positions are indexed from 0 (as in the code in Section D, which uses values 0..n−1), but the theorem and proof never state this convention. Restate Theorem 4.3 with one consistent indexing scheme; with 1-based indices the right-hand side should be (k−1)−L(X^{-1})_k.
  3. [Section 5.2, Figure 6] The cyclic benchmark in Section 5.2 reports only support and validity metrics: % unique, % unique and valid, and % unique, valid and cyclic, together with the observation that about 20% of samples fall in the training set. These metrics do not test whether the learned distribution matches the target uniform distribution over cyclic permutations. A sampler that always outputs a small set of cyclic permutations whose train/holdout split is 20/80 would pass all three checks. The paper's claim that the model 'learns non-trivial distributions even in the least expressive mode' is therefore not supported by this experiment; please add a proper distributional comparison (e.g., exact log-likelihood under the uniform law for n=10, total variation distance, or a chi-square test over the 362,880 cyclic permutations).
  4. [Section 5.3, Figure 7] Section 5.3 evaluates the MovieLens model with NDCG@k. NDCG@k is a top-weighted ranking-quality metric, not a proper scoring rule for the predicted distribution over rankings; a model can achieve high NDCG@k even if its uncertainty is wrong, as long as the top-k order is mostly correct. Since the paper's central claim is about learning distributions, at least one distributional metric (held-out log-likelihood or a proper scoring rule over rankings) should be reported for the MovieLens experiments, together with variance estimates.
minor comments (4)
  1. [Section 5.1, Figure 5] No error bars or number of seeds are reported, and the state-of-the-art claim does not compare against the Lehmer-code autoregressive baseline of Diallo et al. (2020), which is cited in Section 2; please add this comparison or justify its exclusion.
  2. [Section 3.1] The text 'forms a bijections with the symmetric group' should read 'forms a bijection with the symmetric group'.
  3. [Section 4.2, Remark 4.1] The notation P_{L_j|L_{<j}} is introduced without explicitly stating that L here denotes the right-Lehmer code; clarify the code convention to avoid confusion with the left-Lehmer code used in Theorem 4.3.
  4. [Section 5.2, Figure 6] The text says 'All methods can fully model the target distribution at full NFEs' but no distributional metric is shown; at minimum the plot should include a line for the uniform-law log-likelihood of the empirical samples.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular steps: the bijection and decoding results are proved from definitions, and the only self-citations are non-load-bearing implementation references.

full rationale

The derivation chain is self-contained. Lehmer codes, Fisher-Yates draws, and insertion vectors are defined through standard external algorithms (Lehmer, 1960; Fisher and Yates, 1953; Doignon et al., 2004; Sattolo, 1986), and their bijective correspondence with the symmetric group is an immediate consequence of those definitions, not of a fitted quantity. Theorem 4.3 is proved directly in Section B.2 from the definitions of insertion vectors, inverse permutations, and left-Lehmer codes via a change of variables; the equality V(X)_k = k - L(X^{-1})_k is derived, not assumed. The Mallows and RIM containment claims (Remarks 4.1 and 4.2) are explicit specializations of the conditional distributions, so they are containment statements rather than predictions inferred from data. The only self-citations of note are to Kunze et al. (2024a,b) for Lehmer/Fisher-Yates encoding and decoding; although an author of the present paper overlaps, those references are implementation-level and the algorithms are restated as runnable code in Appendices D.1-D.3, so they do not carry the central argument. In-scope limitation statements were weighed: Section 6 concedes 'there is room for more characterization of the properties of these families of distributions in the low NFE setting', and Section C notes that n forward passes are needed for full expressivity; these are acknowledged gaps. The skeptic's point that the cyclic benchmark's % Unique / % Valid / % Cyclic metrics and the MovieLens NDCG metric do not directly measure distributional fidelity is a validation weakness, not circularity, because the model's construction and Theorem 4.3 do not rely on those measurements. No prediction in the paper is equivalent by construction to a fitted input, and no load-bearing step resolves to a self-citation chain.

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

The central claim rests on standard combinatorial bijections and on empirical assumptions about the learning objective and benchmark tasks. No free parameters are fitted post hoc, and no new entities are introduced.

assumptions (5)
  • standard math Lehmer codes, Fisher-Yates draws, and insertion vectors are bijections with the symmetric group.
    Used in Section 3.1 to define the three representations and to guarantee that any sequence within the stated supports decodes to a valid permutation.
  • standard math The chain rule over a bijective representation can express any probability distribution over permutations at full autoregressive decoding.
    Invoked in Section 4.2 to argue that the factorized representations admit arbitrary distributions when the number of forward passes equals the permutation length.
  • standard math Sattolo's algorithm, with Fisher-Yates draws restricted to positive values, generates exactly the cyclic permutations of length n.
    Used in Sections 3.1 and 5.2 to define the cyclic benchmark and to explain why positive draws guarantee cyclic outputs.
  • domain assumption Cross-entropy training on the code-token conditionals is an effective surrogate for learning the target distribution over permutations.
    Section 4.2 adopts the standard language-modeling objective without a derivation that it minimizes a valid statistical divergence to the target permutation distribution in the low-NFE regime.
  • domain assumption The three benchmark tasks (CIFAR-10 jigsaw, uniform cyclic permutations, MovieLens re-ranking) are representative of permutation-learning problems with practical value.
    Section 5 uses these tasks to support the generality of the method; the representativeness is argued but not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Distributions over Permutations and Rankings with Factorized Representations." pith.science (2026). https://pith.science/paper/JRF3I4EW

@misc{pith2026250524664,
  author       = {Pith},
  title        = {Pith review of: Learning Distributions over Permutations and Rankings with Factorized Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JRF3I4EW}},
  note         = {Machine review of arXiv:2505.24664}
}
read the original abstract

Learning distributions over permutations is a fundamental problem in machine learning, with applications in ranking, combinatorial optimization, structured prediction, and data association. Existing methods rely on mixtures of parametric families or neural networks with expensive variational inference procedures. In this work, we propose a novel approach that leverages alternative representations for permutations, including Lehmer codes, Fisher-Yates draws, and Insertion-Vectors. These representations form a bijection with the symmetric group, allowing for unconstrained learning using conventional deep learning techniques, and can represent any probability distribution over permutations. Our approach enables a trade-off between expressivity of the model family and computational requirements. In the least expressive and most computationally efficient case, our method subsumes previous families of well established probabilistic models over permutations, including Mallow's and the Repeated Insertion Model. Experiments indicate our method significantly outperforms current approaches on the jigsaw puzzle benchmark, a common task for permutation learning. However, we argue this benchmark is limited in its ability to assess learning probability distributions, as the target is a delta distribution (i.e., a single correct solution exists). We therefore propose two additional benchmarks: learning cyclic permutations and re-ranking movies based on user preference. We show that our method learns non-trivial distributions even in the least expressive mode, while traditional models fail to even generate valid permutations in this setting.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 12 canonical work pages

  1. [1]

    Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg

    Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces, 2021. https://arxiv.org/abs/2107.03006

    arXiv 2021

  2. [2]

    From ranknet to lambdarank to lambdamart: An overview

    Christopher JC Burges. From ranknet to lambdarank to lambdamart: An overview. Learning, 11 0 (23-581): 0 81, 2010

    2010

  3. [3]

    Elements of information theory

    Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999

    1999

  4. [4]

    Differentiable Ranks and Sorting using Optimal Transport

    Marco Cuturi, Olivier Teboul, and Jean-Philippe Vert. Differentiable ranks and sorting using optimal transport, 2019. https://arxiv.org/abs/1905.11885

    work page Pith review arXiv 2019

  5. [5]

    Bert: Pre-training of deep bidirectional transformers for language understanding, 2019

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. https://arxiv.org/abs/1810.04805

    arXiv 2019

  6. [6]

    ssatou Diallo, Markus Zopf, and Johannes F \

    A \" ssatou Diallo, Markus Zopf, and Johannes F \"u rnkranz. Permutation learning via lehmer codes. In ECAI 2020, pages 1095--1102. IOS Press, 2020

    work page 2020

  7. [7]

    The repeated insertion model for rankings: Missing link between two subset choice models

    Jean-Paul Doignon, Aleksandar Peke c , and Michel Regenwetter. The repeated insertion model for rankings: Missing link between two subset choice models. Psychometrika, 69 0 (1): 0 33--54, 2004

    work page 2004

  8. [8]

    Revisit recommender system in the permutation prospective

    Yufei Feng, Yu Gong, Fei Sun, Junfeng Ge, and Wenwu Ou. Revisit recommender system in the permutation prospective. arXiv preprint arXiv:2102.12057, 2021

    arXiv 2021

Show all 39 references
  1. [9]

    Statistical tables for biological, agricultural and medical research

    Ronald Aylmer Fisher and Frank Yates. Statistical tables for biological, agricultural and medical research. Hafner Publishing Company, 1953

    1953

  2. [10]

    E. N. Gilbert. Theory of shuffling. Technical memorandum, Bell Laboratories, 1955

    1955

  3. [11]

    Stochastic optimization of sorting networks via continuous relaxations, 2019

    Aditya Grover, Eric Wang, Aaron Zweig, and Stefano Ermon. Stochastic optimization of sorting networks via continuous relaxations, 2019. https://arxiv.org/abs/1903.08850

    2019 arXiv

  4. [12]

    The movielens datasets: History and context

    F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis), 5 0 (4): 0 1--19, 2015

    2015

  5. [13]

    Diffusionbert: Improving generative masked language models with diffusion models, 2022

    Zhengfu He, Tianxiang Sun, Kuanning Wang, Xuanjing Huang, and Xipeng Qiu. Diffusionbert: Improving generative masked language models with diffusion models, 2022. https://arxiv.org/abs/2211.15029

    2022 arXiv

  6. [14]

    a rvelin and Jaana Kek \

    Kalervo J \"a rvelin and Jaana Kek \"a l \"a inen. Cumulated gain-based evaluation of ir techniques. ACM Transactions on Information Systems (TOIS), 20 0 (4): 0 422--446, 2002

    2002

  7. [15]

    Jorge K. S. Kamassury, Henrique Pickler, Filipe R. Cordeiro, and Danilo Silva. Cct: A cyclic co-teaching approach to train deep neural networks with noisy labels. IEEE Access, 13: 0 43843--43860, 2025. doi:10.1109/ACCESS.2025.3548510

    2025

  8. [16]

    A new measure of rank correlation

    Maurice G Kendall. A new measure of rank correlation. Biometrika, 30 0 (1-2): 0 81--93, 1938

    1938

  9. [17]

    Generalized neural sorting networks with error-free differentiable swap functions, 2024

    Jungtaek Kim, Jeongbeen Yoon, and Minsu Cho. Generalized neural sorting networks with error-free differentiable swap functions, 2024. https://arxiv.org/abs/2310.07174

    2024 arXiv

  10. [18]

    Disk: A diffusion model for structured knowledge, 2024

    Ouail Kitouni, Niklas Nolte, James Hensman, and Bhaskar Mitra. Disk: A diffusion model for structured knowledge, 2024. https://arxiv.org/abs/2312.05253

    2024 arXiv

  11. [19]

    Practical shuffle coding

    Julius Kunze, Daniel Severo, Jan-Willem van de Meent, and James Townsend. Practical shuffle coding. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024 a . https://openreview.net/forum?id=m2DaXpCoIi

    2024

  12. [20]

    Entropy coding of unordered data structures

    Julius Kunze, Daniel Severo, Giulio Zani, Jan-Willem van de Meent, and James Townsend. Entropy coding of unordered data structures. In The Twelfth International Conference on Learning Representations, 2024 b . https://openreview.net/forum?id=afQuNt3Ruh

    2024

  13. [21]

    The neural autoregressive distribution estimator

    Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 29--37. JMLR Workshop and Conference Proceedings, 2011

    2011

  14. [22]

    Teaching combinatorial tricks to a computer

    Derrick H Lehmer. Teaching combinatorial tricks to a computer. Combinatorial Analysis, pages 179--193, 1960

    1960

  15. [23]

    Effective sampling and learning for mallows models with pairwise-preference data

    Tyler Lu and Craig Boutilier. Effective sampling and learning for mallows models with pairwise-preference data. J. Mach. Learn. Res., 15 0 (1): 0 3783--3829, 2014

    2014

  16. [24]

    Individual choice behavior, volume 4

    R Duncan Luce et al. Individual choice behavior, volume 4. Wiley New York, 1959

    1959

  17. [25]

    Non-null ranking models

    Colin L Mallows. Non-null ranking models. i. Biometrika, 44 0 (1/2): 0 114--130, 1957

    1957

  18. [26]

    Learning latent permutations with gumbel-sinkhorn networks

    Gonzalo Mena, David Belanger, Scott Linderman, and Jasper Snoek. Learning latent permutations with gumbel-sinkhorn networks. arXiv preprint arXiv:1802.08665, 2018

    2018 arXiv

  19. [27]

    The llama 3 herd of models, 2024

    Meta. The llama 3 herd of models, 2024. https://arxiv.org/abs/2407.21783

    2024 arXiv

  20. [28]

    Large language diffusion models, 2025

    Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Large language diffusion models, 2025. https://arxiv.org/abs/2502.09992

    2025 arXiv

  21. [29]

    Monotonic differentiable sorting networks, 2022

    Felix Petersen, Christian Borgelt, Hilde Kuehne, and Oliver Deussen. Monotonic differentiable sorting networks, 2022. https://arxiv.org/abs/2203.09630

    2022 arXiv

  22. [30]

    The analysis of permutations

    Robin L Plackett. The analysis of permutations. Journal of the Royal Statistical Society Series C: Applied Statistics, 24 0 (2): 0 193--202, 1975

    1975

  23. [31]

    Language models are unsupervised multitask learners

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019

    2019

  24. [32]

    Simple and effective masked diffusion language models, 2024

    Subham Sekhar Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin T Chiu, Alexander Rush, and Volodymyr Kuleshov. Simple and effective masked diffusion language models, 2024. https://arxiv.org/abs/2406.07524

    2024 arXiv

  25. [33]

    An algorithm to generate a random cyclic permutation

    Sandra Sattolo. An algorithm to generate a random cyclic permutation. Information processing letters, 22 0 (6): 0 315--317, 1986

    1986

  26. [34]

    A mathematical theory of communication

    Claude E Shannon. A mathematical theory of communication. The Bell system technical journal, 27 0 (3): 0 379--423, 1948

    1948

  27. [35]

    Jiaxin Shi, Kehang Han, Zhe Wang, Arnaud Doucet, and Michalis K. Titsias. Simplified and generalized masked diffusion for discrete data, 2025. https://arxiv.org/abs/2406.04329

    2025 arXiv

  28. [36]

    Neural autoregressive distribution estimation

    Benigno Uria, Marc-Alexandre C \^o t \'e , Karol Gregor, Iain Murray, and Hugo Larochelle. Neural autoregressive distribution estimation. Journal of Machine Learning Research, 17 0 (205): 0 1--37, 2016

    2016

  29. [37]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. CoRR, abs/1706.03762, 2017. http://arxiv.org/abs/1706.03762

    2017 arXiv

  30. [38]

    Symmetricdiffusers: Learning discrete diffusion on finite symmetric groups

    Yongxing Zhang, Donglin Yang, and Renjie Liao. Symmetricdiffusers: Learning discrete diffusion on finite symmetric groups. arXiv preprint arXiv:2410.02942, 2024

    2024 arXiv

  31. [39]

    write newline

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

Pith tools

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