Pith. sign in

REVIEW 3 major objections 25 references

Exogenous Dropout: A Simple, Strong Baseline for Corruption-Robust Time Series Forecasting with Covariates

T0 review · 3 major / 0 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read A one-line training trick makes time-series forecasters robust to corrupted covariates without specialized architecture.

desk verdict Solid empirical paper: whole-channel exogenous dropout is a real, transferable robustness baseline, and the BoundEx foil is beaten cleanly; the general “bounds are unnecessary” claim is a bit stronger than the foil supports. read the letter →

arxiv 2607.05452 v1 pith:CWVZZSCN submitted 2026-07-05 cs.LG

Hao Hu , Xue-shan Ai This is my paper
classification cs.LG
keywords exogenousdropouttimeseriesforecastingcovariatescorruptionrobustnessboundedmodulationFiLMgatingendogenousfloor
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

Time-series models that use external covariates often collapse when those feeds are noisy, shifted in time, or missing. This paper asks whether defending against that failure needs elaborate gated or bounded architectures, or whether a simple training intervention is enough. The authors introduce exogenous dropout: during training, whole exogenous channels are randomly zeroed so the model learns not to over-rely on any single covariate. Across electricity prices, reservoir hydrology, and weather, the same intervention improves every architecture tested under noise, misalignment, and fully missing channels while leaving clean accuracy almost unchanged. When applied to an existing dual-correlation network it becomes the most robust model in the study, beating a purpose-built bounded foil that can fall back to an endogenous-only path. The takeaway is practical: apply the dropout to the model you already have before designing a specialized robust architecture.

What carries the argument

Exogenous dropout: a model-agnostic, train-only augmentation that independently zeros entire exogenous channels (historical and future-known) with probability p and rescales survivors, exposing the forecaster to varying covariate availability so corrupted test-time inputs resemble regimes already seen.

What would settle it

Find or design a bounded exogenous architecture that, under the same shared protocol and dropout rates, systematically beats the best unbounded model plus exogenous dropout on clean accuracy and all three corruption types across the released benchmark domains.

Watch

Extended reading notes

Core claim

Architectural boundedness is not required for robustness to corrupted exogenous covariates. Randomly zeroing whole exogenous channels at training time (exogenous dropout) restores robustness under Gaussian noise, temporal misalignment, and fully missing channels for every architecture tested, and an unbounded dual-correlation model trained this way is more robust, in every domain, than a deliberately strong bounded architecture that can provably revert to the endogenous backbone.

Load-bearing premise

The paper treats its purpose-built gated-FiLM model with a fallback residual as a fair stand-in for the whole class of architectural bounds, so that beating that foil is enough to say explicit bounds are unnecessary.

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

3 major / 0 minor

Summary. The paper argues that robustness of exogenous-covariate time series forecasters to noised, temporally misaligned, or fully missing covariates does not require specialized bounded architectures, and can instead be obtained by a one-line, model-agnostic training intervention: whole-channel exogenous dropout (p=0.3, inverted, train-only). On a multi-domain corruption benchmark (EPF five markets, Rapel hydrology, Jena Weather), the intervention improves Gaussian, misalignment, and missing-channel robustness for every architecture tested while preserving clean accuracy. Applied to DAG, it yields the most robust configuration, outperforming BoundEx—a purpose-built gated FiLM foil with a fallback residual that reverts to the endogenous backbone when the gate closes—on clean accuracy and all three corruptions in every domain. Architecture×dropout tables, a two-factor BoundEx ablation, gate diagnostics, a dropout-rate sweep, and a proportional prediction bound (Appendix A) are used to argue that architectural boundedness is not necessary for the observed robustness.

Significance. If the empirical pattern holds, the paper supplies a simple, strong, and immediately usable baseline that future architectural robustness claims in exogenous forecasting should report against, plus a reusable corruption protocol spanning three domains and three corruption types. Strengths include matched dropout on/off across five models, five seeds, per-market head-to-heads, a two-factor ablation of fallback residual vs. dropout, gate measurements under corruption, a dropout-rate sweep, and an explicit (if loose) representation/prediction bound that the authors themselves treat as non-operative. The practical recommendation—apply exogenous dropout before inventing a specialized robust architecture—is actionable and well supported by the tables for the models and domains studied.

major comments (3)
  1. The central negative claim—that architectural boundedness is not necessary for exogenous-corruption robustness (§1 contributions 3–4; §4.2–§4.4; §5; Conclusion)—rests on BoundEx being a sufficiently strong stand-in for the class of bounded designs. Table 6 shows mean gate ḡ stays in ~0.35–0.75 under clean and all corruptions (including fully missing channels), so the fallback residual is never strongly exercised; Appendix A states the prediction bound is only proportional (∥ŷ−ŷfloor∥₂ ≤ ḡ·s·σmax(W)·∥hmod−hendog∥₂) and empirically loose because γ,β are unconstrained and σmax(W)≈0.8–1.4. Beating this foil therefore does not by itself license the general necessity claim. Either (i) strengthen BoundEx (absolute-bounded FiLM, a gate that actually closes under corruption, or a public bounded SOTA such as FTimeXer-style designs once code is available) and re-run the head-to-head, or (ii) restat
  2. Appendix B reports that CrossLinear and iTransformer (Weather) diverge under the shared no-per-dataset-tuning protocol, with divergent seeds retained in the means (clean MSE >2× floor). Table 2 and Figure 2–3 therefore mix stable and unstable baselines. For the architecture×dropout claim, either exclude divergent seeds with a pre-registered rule and report both, or restrict the primary comparison to the stable models (TimeXer, DAG, BoundEx, PatchTST floor) and move the unstable numbers to an appendix with clear labeling. As written, the headline that dropout helps “every architecture” is partly confounded by training instability.
  3. The corruption suite (Gaussian σ=train-channel std, random shift k∈{3…min(24,T)}, all-channel zeroing) is useful but narrow relative to the deployment claims in the Introduction (revised day-ahead prices, forecast drift, pipeline errors). In particular, Missing is the train-time operator taken to the all-channel extreme, so success under Missing is partly in-distribution for dropout models by construction (§3.5, §4.1). The paper should either add at least one out-of-distribution or adversarial exogenous corruption (as the authors themselves flag in Future work), or explicitly scope the robustness claim to the three operators studied and avoid language that implies general deployment robustness.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical robustness claims and the necessity conclusion rest on held-out corruption measurements, not on definitional reductions or self-citation chains.

full rationale

The paper's load-bearing claims are empirical architecture imes dropout comparisons (Tables 2–5) of clean MSE and vs-Floor degradation under three held-out corruption operators, scored against an endogenous-only PatchTST floor. These quantities are measured on chronological test splits with five seeds; they do not reduce by construction to any fitted target or training objective. The representation/prediction bound (Eq. 6 and Proposition 1 / Appendix A) is the standard gated-residual identity ||hout−hendog||=ḡ·||hmod−hendog|| (and its linear-head consequence); the paper states it holds by construction, then immediately shows via Table 6 that ḡ stays in ~0.35–0.75 under all conditions (including fully missing channels) so the bound is empirically loose and is not the operative robustness mechanism. BoundEx is introduced as a controlled foil, not as a uniqueness theorem or self-cited prior result; no author-overlapping citations supply the central premise. There is therefore no self-definitional loop, no fitted-input-called-prediction, and no load-bearing self-citation. The skeptic concern that BoundEx may be an insufficiently strong foil for the class of bounded designs is a generalization/strength-of-evidence issue, not circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

This is an empirical ML methods paper. Load-bearing content is experimental protocol and modeling choices, not a long axiom chain. Free parameters are training/design knobs (especially dropout rate and BoundEx gate/modulation design). Domain assumptions define the corruption benchmark and the endogenous floor. Invented entities are the named intervention and the BoundEx foil; neither is a physical particle-like postulate, but BoundEx is a new architecture introduced to test the necessity claim.

free parameters (5)
  • exogenous_dropout_rate_p = 0.3 (main); sweep 0–0.7
    Primary intervention strength; fixed at 0.3 for main tables, then swept in {0,0.1,0.3,0.5,0.7}. Headline robustness depends on this choice, though the sweep shows a forgiving range for DAG.
  • BoundEx_gate_temperature_and_MLP_init = near-zero FiLM init; τ in gate MLP
    Gate uses temperature-scaled sigmoid and near-zero FiLM MLP init; these shape whether the bound is exercised. Not data-fitted to the robustness metric, but design knobs of the foil.
  • gate_sparsity_penalty_lambda_g = 0 (reported)
    Optional L1 on gate; swept then set to 0 after little effect. Shows robustness claims are not driven by this penalty.
  • shared_training_hyperparameters = L/H=168/24; lr=1e-4
    L=168, H=24, 50 epochs, patience 5, Adam lr 1e-4, MSE; no per-domain retuning. Central comparisons assume this shared protocol is fair.
  • Gaussian_noise_severity_sigma = σ = train-channel std
    Set to per-channel training std (SNR≈1). Defines one of the three corruption axes used to claim robustness.
assumptions (5)
  • domain assumption Endogenous-only PatchTST MSE is the correct floor against which exogenous models should be judged under corruption.
    Introduced in §2.1 and used for all vs-Floor metrics in §4; if a different floor (e.g., another endogenous backbone) is more appropriate, relative rankings can shift.
  • domain assumption Gaussian noise, random temporal misalignment, and full channel zeroing adequately represent deployment covariate failures for the claim.
    Corruption protocol in §4.1; limitations §6 note adversarial/OOD regimes are future work.
  • domain assumption Supplying recorded future exogenous values as perfect forecasts on Rapel/Weather is a conservative clean-condition setup.
    Stated in §4.1; if perfect future covariates overstate clean gains or change dropout effects, domain transfer of the claim weakens.
  • standard math For a linear forecast head, ||ŷ−ŷ_floor||₂ ≤ ḡ·s·σ_max(W)·||h_mod−h_endog||₂ (Proposition 1).
    Appendix A; standard operator-norm argument on a gated residual. Used to define BoundEx’s proportional guarantee, not the empirical winner.
  • ad hoc to paper Whole-channel inverted dropout during training is a fair, architecture-agnostic intervention when applied identically to all models.
    §3.5 definition and matched application in §4; the paper’s model-agnostic claim rests on this operationalization.
invented entities (3)
  • Exogenous dropout (whole-channel train-time mask) independent evidence
    purpose: Model-agnostic training intervention claimed to deliver corruption robustness without architectural bounds.
    Named and isolated as the recommended baseline; related to prior stochastic exogenous masking but treated as the paper’s primary method.
  • BoundEx (bounded gated-FiLM exogenous modulation with fallback residual)
    purpose: Deliberately strong architectural foil to test whether explicit boundedness is necessary for robustness.
    Author-constructed architecture (§3.1–3.4); independent evidence is internal ablations and gate diagnostics, not external prior validation.
  • Exogenous-corruption robustness benchmark (EPF/Rapel/Weather protocol)
    purpose: Standardized evaluation of clean vs Gaussian/misalign/missing degradations relative to endogenous floor.
    Released contribution claimed in abstract/conclusion; usefulness depends on community adoption beyond this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exogenous Dropout: A Simple, Strong Baseline for Corruption-Robust Time Series Forecasting with Covariates." pith.science (2026). https://pith.science/paper/CWVZZSCN

@misc{pith2026260705452,
  author       = {Pith},
  title        = {Pith review of: Exogenous Dropout: A Simple, Strong Baseline for Corruption-Robust Time Series Forecasting with Covariates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWVZZSCN}},
  note         = {Machine review of arXiv:2607.05452}
}
read the original abstract

Time series forecasters that use exogenous covariates are fragile in deployment: when those covariates are noised, temporally misaligned, or missing, strong exogenous-fusion and exogenous-adapted models can degrade far above the endogenous-only floor. We study whether such robustness requires specialized architectures, or whether it can be obtained through a simple training intervention. We propose exogenous dropout, a model-agnostic method that randomly zeros whole exogenous channels during training. Across electricity-price forecasting, reservoir hydrology, and meteorology, exogenous dropout substantially improves robustness under Gaussian noise, temporal misalignment, and fully missing channels, while preserving clean accuracy. Applied to a dual-correlation network, it yields the most robust model in our experiments, outperforming a deliberately strong bounded architectural foil, BoundEx, which combines a learnable gate, a fallback residual to the endogenous backbone, and per-channel exogenous FiLM modulation. Architecture-by-dropout ablations, gate-behavior diagnostics, and a representation-level bound show that explicit architectural boundedness is not necessary for this robustness: an unbounded model trained with exogenous dropout is more robust than the bounded model in every domain. We release a corruption-robustness benchmark and recommend exogenous dropout as a simple, strong baseline for future work on time series forecasting with covariates.

Figures

Figures reproduced from arXiv: 2607.05452 by the authors.

Figure 1
Figure 1. Overview of the BoundEx bounded gated-FiLM modulation mechanism. The endogenous [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Effect of exogenous dropout on vs-Floor degradation (%; lower is better; light [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Accuracy–robustness trade-off, one panel per domain (absolute clean MSE is comparable only [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Exogenous-dropout rate (p) sensitivity across all three domains (columns) under Gaussian noise (top), misalignment (middle), and missing channels (bottom), for DAG (the winner) and BoundEx (the bounded foil); y-axes are vs-Floor degradation and are scaled per panel, wi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages

  1. [1]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Caner Türkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Olek- sandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschieg- ner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Wilson, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: Learning the language of ...

    arXiv

  2. [2]

    URLhttps://doi.org/10.48550/arxiv

    doi: 10.48550/arxiv.2403.07815. URLhttps://doi.org/10.48550/arxiv. 2403.07815. Abhimanyu Das, Weihao Kong, Andrew Leach, Shaan Mathur, Rajat Sen, and Rose Yu. Long-term fore- casting with tide: Time-series dense encoder.arXiv preprint arXiv:2304.08424,

    arXiv doi:10.48550/arxiv.2403.07815

  3. [3]

    2304.08424

    doi: 10.48550/arxiv. 2304.08424. URLhttps://doi.org/10.48550/arxiv.2304.08424. Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated con- volutional networks. InInternational Conference on Machine Learning (ICML),

    arXiv doi:10.48550/arxiv

  4. [4]

    Vijay Ekambaram, Arindam Jati, Nam Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    URLhttps: //arxiv.org/abs/1612.08083. Vijay Ekambaram, Arindam Jati, Nam Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. Tsmixer: Lightweight mlp-mixer model for multivariate time series forecasting

    arXiv

  5. [5]

    URLhttps://doi.org/10.1145/3580305.3599533

    doi: 10.1145/3580305.3599533. URLhttps://doi.org/10.1145/3580305.3599533. Vaswani et al. Attention is all you need. InNeurIPS,

    doi:10.1145/3580305.3599533

  6. [6]

    URLhttps://doi.org/10.1162/neco.1997.9.8.1735

    doi: 10.1162/neco.1997.9.8.1735. URLhttps://doi.org/10.1162/neco.1997.9.8.1735. Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. Reversible instance normalization for accurate time-series forecasting against distribution shift. InInternational Conference on Learning Representations (ICLR),

    doi:10.1162/neco.1997.9.8.1735 1997

  7. [8]

    Preprint, not yet peer-reviewed

    URLhttps://arxiv.org/abs/2604.02347. Preprint, not yet peer-reviewed. Closest prior work to BoundEx: FFT frequency branch + gated time-frequency fusion + stochastic ex- ogenous masking + consistency regularization for empirical spurious-correlation suppression. Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qings...

    arXiv

  8. [9]

    doi: 10.1145/3637528. 3671451. URLhttps://doi.org/10.1145/3637528.3671451. Bryan Lim, Sercan Ö. Arık, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for interpretable multi-horizon time series forecasting.International Journal of Forecasting,

    doi:10.1145/3637528

Show all 25 references
  1. [10]

    2021.03.012

    doi: 10.1016/j.ijforecast. 2021.03.012. URLhttps://doi.org/10.1016/j.ijforecast.2021.03.012. YongLiu, TenggeHu, HaoranZhang, HaixuWu, ShiyuWang, LintaoMa, andMingshengLong. itransformer: Inverted transformers are effective for time series forecasting.arXiv preprint arXiv:2310.06625,

    2021 doi arXiv

  2. [11]

    URLhttps://doi.org/10.48550/arxiv.2310.06625

    doi: 10.48550/arxiv.2310.06625. URLhttps://doi.org/10.48550/arxiv.2310.06625. Jiecheng Lu, Xu Han, Yan Sun, and Shihao Yang. Cats: Enhancing multivariate time series forecasting by constructing auxiliary time series as exogenous variables. InInternational Conference on Machine...

    doi arXiv

  3. [12]

    URLhttps://www.semanticscholar.org/paper/ 09e098f5bc0a187bea27437a78a0e9ae83cf6a3d

    doi: 10.48550/arXiv.2403.01673. URLhttps://www.semanticscholar.org/paper/ 09e098f5bc0a187bea27437a78a0e9ae83cf6a3d. Yuqi Nie, Nam Hoai Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arX...

    doi

  4. [13]

    URLhttps://doi.org/10.48550/arxiv.2211.14730

    doi: 10.48550/ arxiv.2211.14730. URLhttps://doi.org/10.48550/arxiv.2211.14730. Kin G. Olivares, Cristian Challu, Grzegorz Marcjasz, Rafał Weron, and Artur Dubrawski. Neural basis ex- pansion analysis with exogenous variables: Forecasting electricity prices with nbeatsx.arXiv p...

    doi arXiv

  5. [14]

    URLhttps://arxiv.org/abs/2104

    doi: 10.1016/j.ijforecast.2022.03.001. URLhttps://arxiv.org/abs/2104. 05522. Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI Conference on Artificial Intelligence,

    2022 doi

  6. [15]

    URLhttps://doi.org/10.1609/aaai.v32i1.11671

    doi: 10.1609/aaai.v32i1.11671. URLhttps://doi.org/10.1609/aaai.v32i1.11671. Xiangfei Qiu, Yuhan Zhu, Zhengyu Li, Xingjian Wu, Bin Yang, and Jilin Hu. Dag: A dual correlation network for time series forecasting with exogenous variables.arXiv preprint arXiv:2509.14933,

    doi arXiv

  7. [16]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean

    URL https://arxiv.org/abs/2509.14933. Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. InInternational Conference on Learning Representation...

    arXiv

  8. [17]

    Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber

    URLhttps://arxiv.org/abs/1701.06538. Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Training very deep networks. InAdvances in Neural Information Processing Systems (NeurIPS),

    arXiv

  9. [18]

    Kshitij Tayal, Arvind Renganathan, Xiaowei Jia, Vipin Kumar, and Dan Lu

    URLhttps://arxiv.org/abs/1507.06228. Kshitij Tayal, Arvind Renganathan, Xiaowei Jia, Vipin Kumar, and Dan Lu. Exotst: Exogenous-aware temporal sequence transformer for time series prediction

    arXiv

  10. [19]

    URL https://doi.org/10.1109/icdm59182.2024.00105

    doi: 10.1109/icdm59182.2024.00105. URL https://doi.org/10.1109/icdm59182.2024.00105. Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Yunzhong Qiu, Haoran Zhang, Jianmin Wang, and Mingsheng Long. Timexer: Empowering transformers for time series forecasting with exogenous variab...

    2024 doi

  11. [20]

    URLhttps://www

    doi: 10.48550/arXiv.2402.19072. URLhttps://www. semanticscholar.org/paper/db5cb5354d8c1ff9cccf2a1c9ef5c218c9323cc1. Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting.arXiv prepri...

    doi arXiv

  12. [21]

    Yosuke Yamaguchi, Issei Suemitsu, and Wei Wu

    URLhttps: //arxiv.org/abs/2106.13008. Yosuke Yamaguchi, Issei Suemitsu, and Wei Wu. Citras: Covariate-informed transformer for time series forecasting.IEEE Access,

    arXiv

  13. [22]

    URLhttps://doi.org/10.1109/ access.2026.3695717

    doi: 10.1109/access.2026.3695717. URLhttps://doi.org/10.1109/ access.2026.3695717. Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? In Proceedings of the AAAI Conference on Artificial Intelligence,

    2026 doi

  14. [23]

    URL https://doi.org/10.1609/aaai.v37i9.26317

    doi: 10.1609/aaai.v37i9.26317. URL https://doi.org/10.1609/aaai.v37i9.26317. Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. InProceedings of the AA...

    doi

  15. [24]

    URLhttps: //doi.org/10.1609/aaai.v35i12.17325

    doi: 10.1609/aaai.v35i12.17325. URLhttps: //doi.org/10.1609/aaai.v35i12.17325. Pengfei Zhou, Yunlong Liu, Junli Liang, Qi Song, and Xiang-Yang Li. Crosslinear: Plug-and-play cross- correlation embedding for time series forecasting with exogenous variables

    doi

  16. [25]

    doi: 10.1145/3711896. 3736899. URLhttps://doi.org/10.1145/3711896.3736899. Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting.arXiv preprint arXiv:2201.12740,

    doi

  17. [26]

    URL https://arxiv.org/abs/2201.12740. 21

    arXiv

Pith tools

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