pith. sign in

arxiv: 2606.03926 · v1 · pith:CJ5J7JKGnew · submitted 2026-06-02 · 💻 cs.HC · cs.LG

DiffUNet²: Bidirectional Prediction, Probabilistic Generation and Collaborative Visual Discovery for Scientific Data

Mengdi Chu , Jiaxin Yang , Angus G. Forbes , Nathan Debardeleben , Earl Lawrence , Ayan Biswas , Han-Wei Shen This is my paper

Pith reviewed 2026-06-28 08:09 UTC · model grok-4.3

classification 💻 cs.HC cs.LG
keywords diffusion modelbidirectional predictionscientific temporal datagenerative modelinginteractive visualizationprobability space navigationany-to-any generationvisual analytics
0
0 comments X

The pith

A conditional diffusion model generates any-to-any bidirectional predictions for scientific time series while modeling multiple plausible outcomes.

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

The paper presents DiffUNet^2, a conditional diffusion model designed to handle temporal evolution in scientific data by supporting predictions in both forward and backward directions from any starting point. It aims to move beyond single deterministic forecasts by capturing full distributions of possible system behaviors. An interactive visual system built on the model lets users explore branching timelines, edit states, and navigate probability spaces. This setup is tested on five datasets from different domains and assessed through expert collaboration to show it can fit into real scientific analysis workflows. If the approach holds, scientists gain a way to treat generative models as active hypothesis-testing tools rather than passive predictors.

Core claim

DiffUNet^2 is a conditional diffusion model that performs bidirectional any-to-any generation across time and captures distributions of plausible system evolutions. The accompanying interactive system supports branching timeline exploration, user-guided state editing, and probability-space navigation, allowing scientists to explore alternative hypotheses in temporal scientific data rather than receiving only deterministic forward predictions.

What carries the argument

DiffUNet^2 conditional diffusion model for bidirectional any-to-any generation across time steps

If this is right

  • The model supports generation of multiple plausible future and past states from any time point in a time series.
  • Users can edit states and observe resulting changes in the probability space of outcomes.
  • Evaluation across five datasets confirms predictive accuracy and quality of the captured ensembles.
  • Expert collaboration shows the combined modeling and visualization approach fits practical scientific temporal analysis tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same bidirectional generation approach could apply to simulation domains not covered in the five evaluated datasets.
  • Interactive navigation of probability space might reduce reliance on post-hoc uncertainty quantification methods in scientific software.
  • This framing suggests generative models can serve as editable hypothesis generators rather than fixed forecasters in data analysis pipelines.

Load-bearing premise

The model accurately captures the true distributions of scientific system evolutions and the interactive features meaningfully support domain experts' workflows.

What would settle it

A test on held-out data where the model's generated probability distributions fail to match observed variability in the scientific systems, or domain experts report that the interactive features do not improve their ability to analyze temporal data.

Figures

Figures reproduced from arXiv: 2606.03926 by Angus G. Forbes, Ayan Biswas, Earl Lawrence, Han-Wei Shen, Jiaxin Yang, Mengdi Chu, Nathan Debardeleben.

Figure 1
Figure 1. Figure 1: Overview of the proposed model and interactive framework for exploring scientific temporal dynamics. (a) The model enables [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Diffusion models capture ambiguity by generating multiple plausi [PITH_FULL_IMAGE:figures/full_fig_p001_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: DiffUNet2 model structure and functionalities. Input temporal data are optionally compressed by a KL-regularized VAE for efficient modeling. DiffUNet2 then performs conditional diffusion generation under the observed state and temporal query, supporting bidirectional any-to-any prediction as well as user-guided editing and refinement through guided denoising. • Accelerating Diffusion through Latent Compres… view at source ↗
Figure 4
Figure 4. Figure 4: User interface of our interactive system. (A,B) Users load data and adjust generation parameters. (C) The node-based trajectory view supports forward generation, backward generation, ensemble generation and branch exploration. (D,E) Users can edit selected states and refine the results with guidance. (F-H) The distribution space view. (I) Users can compare samples of interest to inspect their differences. … view at source ↗
Figure 6
Figure 6. Figure 6: Condition-guided denoising network. The model generates a [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: The diffusion process. The forward process [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Latent compression. High-resolution data are compressed into a [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: User-guided editing and diffusion refinement. Users can directly [PITH_FULL_IMAGE:figures/full_fig_p005_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Temporal conditioning and pair construction for bidirectional any [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Branching temporal exploration interface. (A) Generation control [PITH_FULL_IMAGE:figures/full_fig_p006_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Data distribution space exploration interface. (A) Controls for [PITH_FULL_IMAGE:figures/full_fig_p006_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Ensemble generation results across five datasets: Shallow-water, [PITH_FULL_IMAGE:figures/full_fig_p007_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison of forward prediction across 5 datasets. [PITH_FULL_IMAGE:figures/full_fig_p008_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: HEAT-PLI case study. (A) Bidirectional branching exploration, [PITH_FULL_IMAGE:figures/full_fig_p008_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Wildfire case study. (A) Forward generation from the same [PITH_FULL_IMAGE:figures/full_fig_p009_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: RealPDE-FSI case study. (A) Ensemble generation from the [PITH_FULL_IMAGE:figures/full_fig_p009_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Sequence prediction results on the Shallow-water and Cloverleaf datasets under forward conditioning and backward conditioning. For each [PITH_FULL_IMAGE:figures/full_fig_p012_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: DiffUNet2 can reflect uncertainty by generating multiple plausible outcomes under the same condition. In the HEAT-PLI(density), while the outer-shell evolution remains consistent across ensemble members, differences appear in the later stages, especially in the fine interior shock-wave and material structures. Ens 1 s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 Ens 2 Ens 3 Std 0 5 10 15 20 25 30 0 5 10 15 Ens 1 s1 s2 s3… view at source ↗
Figure 20
Figure 20. Figure 20: Diagnosing ambiguity in Cloverleaf through variable sufficiency. The left panel shows results when only the density field is used, where [PITH_FULL_IMAGE:figures/full_fig_p012_20.png] view at source ↗
read the original abstract

Modeling temporal evolution is important to analyzing and reasoning about scientific phenomena, yet most machine learning methods provide deterministic forward predictions that overlook multiple plausible outcomes and rarely support backward reasoning, limiting their usefulness in practical scientific workflows. We present a framework that integrates diffusion-based generative modeling with interactive visual analytics for scientific exploration. We introduce DiffUNet^2, a conditional diffusion model that enables bidirectional, any-to-any generation across time and captures distributions of plausible system evolutions. Built upon the model, our interactive system supports branching timeline exploration, user-guided state editing, and probability-space navigation, enabling scientists to actively explore alternative hypotheses rather than passively observe predictions. We evaluate the model on 5 datasets across different scientific domains to validate its predictive accuracy and probability-space ensemble quality. In collaboration with domain experts, we demonstrate the effectiveness of our approach in supporting practical scientific temporal data analysis workflows. By integrating modeling and visual interaction, our approach enables scientists to interactively explore system dynamics, transforming generative models into tools for hypothesis-driven scientific analysis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces DiffUNet^2, a conditional diffusion model for bidirectional any-to-any temporal generation across time that captures distributions of plausible system evolutions. It integrates this model with an interactive visual analytics system supporting branching timeline exploration, user-guided state editing, and probability-space navigation. The approach is evaluated on 5 datasets across scientific domains for predictive accuracy and ensemble quality, and demonstrated via collaboration with domain experts for practical scientific workflows.

Significance. If the technical claims are substantiated with appropriate methods and results, the work could meaningfully advance scientific temporal analysis by combining probabilistic generative modeling with interactive tools, allowing hypothesis-driven exploration of multiple plausible evolutions rather than deterministic forward predictions alone.

major comments (1)
  1. [Abstract] Abstract: The manuscript claims evaluation on 5 datasets to validate predictive accuracy and probability-space ensemble quality, yet supplies no methods, architecture details, loss functions, sampling procedures, quantitative results, baselines, error bars, or validation metrics. This absence is load-bearing for the central claim that the model enables effective bidirectional generation and distribution capture.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the major comment below and outline revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The manuscript claims evaluation on 5 datasets to validate predictive accuracy and probability-space ensemble quality, yet supplies no methods, architecture details, loss functions, sampling procedures, quantitative results, baselines, error bars, or validation metrics. This absence is load-bearing for the central claim that the model enables effective bidirectional generation and distribution capture.

    Authors: The abstract is a concise summary; the full manuscript details the DiffUNet^2 architecture (Section 3), bidirectional conditional diffusion formulation with loss functions (Section 3.2), sampling procedures (Section 3.3), and quantitative evaluation across the five datasets including baselines, metrics for accuracy and ensemble quality, and results (Section 4). To address the concern, we will revise the abstract to briefly reference the evaluation metrics and bidirectional generation approach, making the central claims more self-contained while preserving length constraints. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The provided abstract and description introduce DiffUNet^2 as a conditional diffusion model for bidirectional any-to-any temporal generation, with an interactive analytics layer. No equations, fitted parameters, or derivation steps are visible. No self-definitional claims, predictions that reduce to fitted inputs, load-bearing self-citations, uniqueness theorems imported from authors, ansatzes smuggled via citation, or renamings of known results appear. The central claims rest on standard diffusion modeling practices plus evaluation on external datasets and expert collaboration, remaining self-contained without reduction to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no information on free parameters, axioms, or invented entities; all such elements are unknown.

pith-pipeline@v0.9.1-grok · 5730 in / 1110 out tokens · 29678 ms · 2026-06-28T08:09:27.160363+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 32 canonical work pages · 9 internal anchors

  1. [1]

    Andrae, T

    M. Andrae, T. Landelius, J. Oskarsson, and F. Lindsten. Continuous ensemble weather forecasting with diffusion models. InThe Thirteenth International Conference on Learning Representations, 2025. doi: 10. 48550/arXiv.2410.05431 2

    arXiv 2025

  2. [2]

    Anirudh, J

    R. Anirudh, J. J. Thiagarajan, P.-T. Bremer, and B. K. Spears. Improved surrogates in inertial confinement fusion with manifold and cycle consis- tencies.Proceedings of the National Academy of Sciences, 117(18):9741– 9746, 2020. doi: 10.1073/pnas.1916634117 1

    work page doi:10.1073/pnas.1916634117 2020

  3. [3]

    Banesh, S

    D. Banesh, S. Chakrabarti, S. De, G. Egozi, K. Hickmann, W. S. Hlavacek et al. The high explosives & affected targets (heat) dataset. 7 2025. doi: 10.25583/2571471 2, 4, 6

    work page doi:10.25583/2571471 2025

  4. [4]

    Biswas and T

    A. Biswas and T. Turton. Cloverleaf data artifacts for artimis ldrd. 1 2026. doi: 10.25583/3022785 2, 6

    work page doi:10.25583/3022785 2026

  5. [6]

    Casiraghi and G

    G. Casiraghi and G. Andres. Disentangling the timescales of a complex system: A bayesian approach to temporal network analysis, 2024. doi: 10. 48550/arXiv.2403.05343 1

    arXiv 2024

  6. [7]

    Y .-T. Chen, H. Li, N. Shi, X. Luo, W. Xu, and H.-W. Shen. Explorable INR: An Implicit Neural Representation for Ensemble Simulation En- abling Efficient Spatial and Parameter Exploration .IEEE Transactions on Visualization & Computer Graphics, 31(06):3758–3770, June 2025. doi: 10.1109/TVCG.2025.3567052 1

    work page doi:10.1109/tvcg.2025.3567052 2025

  7. [8]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair et al. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Weinberger, eds.,Advances in Neural Information Processing Systems, vol. 27. Curran Associates, Inc., 2014. doi: 10.48550/arXiv.1406.2661 2

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1406.2661 2014

  8. [9]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, 2020. doi: 10.48550/arXiv.2006.11239 2, 4, 9

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2006.11239 2020

  9. [10]

    P. Hu, H. Feng, H. Liu, T. Yan, W. Deng, T. Gao et al. Realpdebench: A benchmark for complex physical systems with real-world data. InThe Fourteenth International Conference on Learning Representations, 2026. Oral Presentation. doi: 10.48550/arXiv.2601.01829 2, 6, 7

    work page doi:10.48550/arxiv.2601.01829 2026

  10. [11]

    Kawakami, D

    Y . Kawakami, D. Cayan, D. Liu, and K.-L. Ma. ClimateSOM: A Visual Analysis Workflow for Climate Ensemble Datasets .IEEE Transactions on Visualization & Computer Graphics, 32(01):473–483, Jan. 2026. doi: 10.1109/TVCG.2025.3634788 2

    work page doi:10.1109/tvcg.2025.3634788 2026

  11. [12]

    D. P. Kingma and M. Welling. An introduction to variational autoencoders. Foundations and Trends in Machine Learning, 12(4):307–392, 11 2019. doi: 10.1561/2200000056 2

    work page doi:10.1561/2200000056 2019

  12. [13]

    R. Lam, A. Sanchez-Gonzalez, M. Willson, P. Wirnsberger, M. Fortunato, F. Alet et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416–1421, 2023. doi: 10.1126/science.adi2336 2

    work page doi:10.1126/science.adi2336 2023

  13. [15]

    L. Li, R. Carver, I. Lopez-Gomez, F. Sha, and J. Anderson. Generative emulation of weather forecast ensembles with diffusion models.Science Advances, 10:eadk4489, 2024. doi: 10.1126/sciadv.adk4489 2, 9

    work page doi:10.1126/sciadv.adk4489 2024

  14. [16]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart et al. Fourier neural operator for parametric partial differential equations,

  15. [17]

    doi: 10.48550/arXiv.2010.08895 2

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2010.08895 2010

  16. [18]

    H. Liu, C. Liu, J. T. L. Wang, and H. Wang. Predicting solar flares using a long short-term memory network.The Astrophysical Journal, 877(2):121, jun 2019. doi: 10.3847/1538-4357/ab1b3c 2

    work page doi:10.3847/1538-4357/ab1b3c 2019

  17. [19]

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G. E. Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3):218–229, Mar. 2021. doi: 10 .1038/s42256-021-00302-5 2

    2021

  18. [20]

    Ma and J

    X. Ma and J. Li. Latentzoom: Seamless scaling in generative latent space for visual exploration of local performance in deep neural networks.IEEE Transactions on Visualization and Computer Graphics, 32(3):2709–2725,

  19. [21]

    doi: 10.1109/TVCG.2026.3654030 2

    work page doi:10.1109/tvcg.2026.3654030 2026

  20. [22]

    Moreland

    K. Moreland. A survey of visualization pipelines.IEEE Transactions on Visualization and Computer Graphics, 19(3):367–378, 2013. doi: 10. 1109/TVCG.2012.133 1

    2013

  21. [23]

    Improved Denoising Diffusion Probabilistic Models

    A. Nichol and P. Dhariwal. Improved denoising diffusion probabilistic models, 2021. doi: 10.48550/arXiv.2102.09672 2

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2102.09672 2021

  22. [24]

    O’Shea and R

    K. O’Shea and R. Nash. An introduction to convolutional neural networks,

  23. [25]

    doi: 10.48550/arXiv.1511.08458 1

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1511.08458

  24. [26]

    v24i1.7740

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI conference on artificial intelligence, vol. 32, 2018. doi: 10.1609/aaai .v32i1.11671 4

    work page doi:10.1609/aaai 2018

  25. [27]

    Choudhary, J

    T. Pfaff, M. Fortunato, A. Sanchez-Gonzalez, and P. Battaglia. Learning mesh-based simulation with graph networks. InInternational Conference on Learning Representations (ICLR), 2021. doi: 10.48550/arXiv.2010. 03409 2

    work page doi:10.48550/arxiv.2010 2021

  26. [28]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707, 2019. doi: 10.1016/j.jcp.2018.10. 045 1

    work page doi:10.1016/j.jcp.2018.10 2019

  27. [29]

    Rühling Cachay, B

    S. Rühling Cachay, B. Zhao, H. Joren, and R. Yu. Dyffusion: A dynamics- informed diffusion model for spatiotemporal forecasting. In A. Oh, T. Nau- mann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, eds.,Advances in Neural Information Processing Systems, vol. 36, pp. 45259–45287. Curran Associates, Inc., 2023. doi: 10.48550/arXiv.2306.01984 2

    work page doi:10.48550/arxiv.2306.01984 2023

  28. [30]

    R., McLeod, A

    A. Sanchez-Gonzalez, J. Godwin, T. Pfaff, R. Ying, J. Leskovec, and P. W. Battaglia. Learning to simulate complex physics with graph networks. In Proceedings of the 37th International Conference on Machine Learning, ICML’20, art. no. 784, 10 pages. JMLR.org, 2020. doi: 10.48550/arXiv. 2002.09405 2

    work page internal anchor Pith review doi:10.48550/arxiv 2020

  29. [31]

    Schneider, J

    T. Schneider, J. Teixeira, C. S. Bretherton, et al. Climate goals and computing the future of clouds.Nature Climate Change, 7(1):3–5, Jan

  30. [32]

    doi: 10.1038/nclimate3190 1

    work page doi:10.1038/nclimate3190

  31. [33]

    Shu and A

    D. Shu and A. B. Farimani. Zero-shot uncertainty quantification using diffusion probabilistic models, 2024. doi: 10.48550/arXiv.2408.04718 2, 9

    work page doi:10.48550/arxiv.2408.04718 2024

  32. [34]

    Q. Shu, H. Guo, J. Liang, L. Che, J. Liu, and X. Yuan. Ensemblegraph: Interactive visual analysis of spatiotemporal behaviors in ensemble simu- lation data. In2016 IEEE Pacific Visualization Symposium (PacificVis), pp. 56–63, 2016. doi: 10.1109/PACIFICVIS.2016.7465251 2

    work page doi:10.1109/pacificvis.2016.7465251 2016

  33. [35]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models,

  34. [36]

    doi: 10.48550/arXiv.2010.02502 2

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2010.02502 2010

  35. [37]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. doi: 10.48550/arXiv.2011.13456 2

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2011.13456 2021

  36. [38]

    S. H. Strogatz.Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering (3rd ed.). Chapman and Hall/CRC., 2024. doi: 10.1201/9780429398490 1

    work page doi:10.1201/9780429398490 2024

  37. [39]

    Takamoto, T

    M. Takamoto, T. Praditia, R. Leiteritz, D. MacKinlay, F. Alesiani, D. Pflüger et al. Pdebench: an extensive benchmark for scientific machine learning. InProceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, art. no. 117, 16 pages. Curran Associates Inc., Red Hook, NY , USA, 2022. doi: 10.48550/arXiv.2210. 07182 2, 6

    work page doi:10.48550/arxiv.2210 2022

  38. [40]

    Attention Is All You Need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez et al. Attention is all you need, 2023. doi: 10.48550/arXiv.1706.03762 1

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1706.03762 2023

  39. [41]

    J. Wang, S. Hazarika, C. Li, and H.-W. Shen. Visualization and visual analysis of ensemble data: A survey.IEEE Transactions on Visualization and Computer Graphics, 25(9):2853–2872, 2019. doi: 10.1109/TVCG. 2018.2853721 2

    work page doi:10.1109/tvcg 2019

  40. [42]

    W. Yu, A. Ghosh, T. S. Finn, R. Arcucci, M. Bocquet, and S. Cheng. A probabilistic approach to wildfire spread prediction using a denoising diffusion surrogate model.Geoscientific Model Development, 19(2):1027– 1054, 2026. doi: 10.5194/gmd-19-1027-2026 2, 6, 7, 9

    work page doi:10.5194/gmd-19-1027-2026 2026

  41. [43]

    A. Zhou, Z. Li, M. Schneier, J. R. B. Jr, and A. B. Farimani. Text2PDE: Latent diffusion models for accessible physics simulation. InThe Thir- teenth International Conference on Learning Representations, 2025. doi: 10.48550/arXiv.2410.01153 2 A APPENDIX This section provides supplementary results and figures that complement the main paper. s=1 s=2 s=3 s=4...

    work page doi:10.48550/arxiv.2410.01153 2025