Cascade-KDE: Robust Time-Series Restoration under Out-of-Distribution Impulse Corruptions

Ning Yang; Yuefeng Liu; Ziyu Yang

arxiv: 2605.24055 · v1 · pith:2XXPMKR2new · submitted 2026-05-22 · 💻 cs.LG · cs.AI

Cascade-KDE: Robust Time-Series Restoration under Out-of-Distribution Impulse Corruptions

Yuefeng Liu , Ning Yang , Ziyu Yang This is my paper

Pith reviewed 2026-06-30 16:00 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords time-series restorationimpulse corruptionkernel density estimationoutlier handlingderivative preservationtraining-free methodsignal preprocessing

0 comments

The pith

Cascade-KDE restores time series by truncating a 2D temporal-amplitude density to exclude impulse outliers while preserving local features.

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

The paper presents Cascade-KDE as a training-free method to restore time series that contain both Gaussian noise and occasional large impulse outliers. It first builds a two-dimensional density over time and amplitude values drawn from the corrupted observations, then applies a Density-Truncated Robust Expectation step that limits the pull of distant abnormal points, and finally runs an exponential cascade of refinements that stops adaptively. The design focuses on keeping derivative peaks and task-critical local shapes intact, which matters for uses such as ECG analysis or battery monitoring where shape drives later decisions. Reported benchmark results show gains over classical filters and learning baselines in reconstruction accuracy, derivative fidelity, downstream classification, and speed.

Core claim

Cascade-KDE estimates a two-dimensional temporal-amplitude density from corrupted observations, applies Density-Truncated Robust Expectation to limit influence of abnormal points, and refines the sequence through an exponential cascade with adaptive stopping, producing restored trajectories that remain close to the original local structure under out-of-distribution impulse corruptions.

What carries the argument

Density-Truncated Robust Expectation applied to a two-dimensional temporal-amplitude density estimate.

If this is right

Restored series achieve higher curve fidelity and better derivative preservation than classical filters or learning-based baselines.
Downstream classification accuracy improves when the restored data is used as input.
Runtime is lower than the compared baselines while still delivering the fidelity gains.
Bounded density-based restoration can serve as a practical preprocessing step in noisy time-series pipelines.

Where Pith is reading between the lines

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

The same truncation logic could be tested on multivariate sensor streams where impulses appear in only some channels.
Adaptive stopping in the cascade might reduce unnecessary iterations in other iterative restoration algorithms.
Controlled experiments that vary the fraction of impulses while holding signal features fixed would directly test the separation claim.

Load-bearing premise

A two-dimensional temporal-amplitude density estimated from the corrupted observations, combined with density truncation, can reliably separate out-of-distribution impulses from the underlying signal without distorting task-critical local features.

What would settle it

A dataset in which genuine high-amplitude signal features occupy low-density regions in the temporal-amplitude plane, so that truncation removes or distorts those features.

Figures

Figures reproduced from arXiv: 2605.24055 by Ning Yang, Yuefeng Liu, Ziyu Yang.

**Figure 1.** Figure 1: Cascade-KDE pipeline overview. The final artwork should summarize the full restoration flow from corrupted input to feature-preserving output. construct a continuous spatial probability density function: pˆ(t, y) = 1 Nhthy X N i=1 K t − ti ht , y − yi hy (2) where K is the standard 2D Gaussian kernel, and ht, hy are the bandwidth parameters. Dense clusters of true signal form a high-density ridge, whil… view at source ↗

**Figure 2.** Figure 2: Pareto-guided adaptive cascade depth selection. where Hsmoothness = std(d 2y/dt2 ) measures the reduction of high-frequency variation, while Fsharpness = max |d 2y/dt2 | tracks the preservation of sharp local structure. 5 Theoretical Analysis 5.1 Sensitivity of Unbounded Conditional Expectation The classical unbounded conditional expectation can be influenced by remote high-amplitude anomalies. If an impul… view at source ↗

**Figure 3.** Figure 3: Comparison of deep learning baselines and Cascade-KDE under unseen out-ofdistribution impulse corruption [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Compact quantitative summary of robustness and downstream utility. 6.5 Impact on Downstream Classification A denoising algorithm is ultimately judged by its utility to downstream machine learning applications. Using the standard ECG5000 dataset via the UCR archive, we applied impulse corruption to the test set and evaluated an SVM-based classification task. The noisy input reduced accuracy to 71.33%. The … view at source ↗

**Figure 5.** Figure 5: Adaptive K∗ search balancing derivative smoothness and sharpness. 6.7 Edge-Deployable Scalability A critical factor for practical use is runtime. We benchmarked the end-to-end processing time of Cascade-KDE across sequence lengths N ∈ [100, 4000]. For typical sliding-window lengths (N = 250 to 500), Cascade-KDE processes a sequence in approximately 125 ms to 429 ms on a standard CPU. This indicates that th… view at source ↗

read the original abstract

Real-world time-series data in industrial sensing, healthcare, and energy systems is often corrupted by a mixture of Gaussian noise and occasional large-magnitude impulse outliers. For tasks that depend on local shape, such as ECG morphology analysis and battery degradation monitoring, the main requirement is not only low reconstruction error but also preservation of derivative peaks and task-critical features. We propose Cascade-KDE, a training-free restoration framework for corrupted time series. The method first estimates a two-dimensional temporal-amplitude density, then applies a Density-Truncated Robust Expectation to limit the influence of distant abnormal points, and finally refines the sequence through an exponential cascade with adaptive stopping. This design aims to improve robustness under out-of-distribution impulse corruptions while keeping the restored trajectory close to the original local structure. Across several benchmark datasets, the proposed method shows consistent gains over classical filters and representative learning-based baselines on curve fidelity, derivative preservation, downstream classification, and runtime efficiency. These results suggest that bounded density-based restoration is a practical option for feature-preserving preprocessing in noisy time-series pipelines.

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.

Desk Editor's Note private letter to a colleague

Cascade-KDE is a training-free KDE-plus-cascade filter for impulse noise whose main separation step rests on an untested assumption about density overlap.

read the letter

The paper's core idea is a training-free pipeline that builds a 2D temporal-amplitude density from the corrupted series itself, truncates low-density points inside a robust expectation step, and then applies an adaptive exponential cascade. That combination is presented as new, and the training-free property is genuinely useful for settings where you cannot collect clean paired data.

What it does reasonably well is name the right practical constraints: derivative preservation and downstream task performance matter more than pure MSE in ECG or sensor work, and the method tries to address both. The abstract also flags runtime efficiency as a selling point, which matters for real pipelines.

The soft spot is the load-bearing claim that density truncation on the corrupted observations will reliably down-weight only the impulses. Because the density is estimated from the mixture, any legitimate high-amplitude excursions that sit near the impulse distribution will either stay corrupted or get smoothed. The abstract supplies no analytic bound, no sensitivity plot on the truncation threshold, and no ablation that isolates the truncation operator, so the reported gains on fidelity and derivatives cannot be checked against that risk. The lack of any numbers, error bars, or dataset details in the provided text makes the empirical part impossible to evaluate.

This is aimed at applied time-series users who need a quick preprocessing step rather than a theoretical advance. A reader who already works with robust KDE variants will find the cascade incremental; someone looking for a drop-in filter might still want to test it. The work shows clear thinking about the problem constraints and avoids the usual circularity traps of learned methods, so it is coherent on its own terms.

I would send it to review so the experiments and any missing derivations can be examined, but the current evidence is too thin to judge the central assumption.

Referee Report

2 major / 0 minor

Summary. The paper proposes Cascade-KDE, a training-free framework for restoring time series corrupted by Gaussian noise and out-of-distribution impulse outliers. It estimates a two-dimensional temporal-amplitude density directly from the corrupted input, applies Density-Truncated Robust Expectation to limit the influence of abnormal points, and refines the sequence via an exponential cascade with adaptive stopping. The central claim is that this yields consistent gains over classical filters and learning-based baselines on curve fidelity, derivative preservation, downstream classification accuracy, and runtime efficiency across several benchmark datasets, while preserving task-critical local structure.

Significance. If the separation assumption holds, the approach could provide a practical training-free preprocessing option for feature-sensitive applications such as ECG morphology analysis and battery degradation monitoring, where both reconstruction accuracy and derivative fidelity matter. The emphasis on bounded density estimation from corrupted observations alone is a potentially useful direction for robust time-series pipelines.

major comments (2)

[Method description (abstract and §3)] The core assumption that a 2D temporal-amplitude density estimated from the corrupted observations, followed by density truncation inside the Robust Expectation step, can systematically isolate OOD impulses without distorting task-critical local features or derivatives is load-bearing for all reported gains. No analytic bound on the truncation threshold, no sensitivity analysis, and no ablation isolating the truncation operator are supplied, leaving the separation claim unverified even if tables appear favorable.
[Abstract and Results] Abstract asserts empirical gains on curve fidelity, derivative preservation, downstream classification, and runtime across multiple benchmark datasets, yet supplies neither quantitative numbers, error bars, dataset descriptions, statistical tests, nor implementation details. This prevents any assessment of whether the claimed improvements are statistically meaningful or reproducible.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the central assumptions of Cascade-KDE and the presentation of empirical results. We address each major comment below, indicating planned revisions where appropriate.

read point-by-point responses

Referee: [Method description (abstract and §3)] The core assumption that a 2D temporal-amplitude density estimated from the corrupted observations, followed by density truncation inside the Robust Expectation step, can systematically isolate OOD impulses without distorting task-critical local features or derivatives is load-bearing for all reported gains. No analytic bound on the truncation threshold, no sensitivity analysis, and no ablation isolating the truncation operator are supplied, leaving the separation claim unverified even if tables appear favorable.

Authors: The referee correctly notes that the separation assumption underpins the reported gains. The Density-Truncated Robust Expectation step is designed so that OOD impulses fall into low-density regions of the estimated 2D temporal-amplitude distribution and are thereby down-weighted, while in-distribution points retain influence; this is motivated by principles from robust kernel density estimation. We do not supply an analytic bound on the truncation threshold. However, we will add a sensitivity analysis across a range of truncation values and an ablation that isolates the truncation operator from the rest of the cascade in the revised manuscript to provide stronger empirical verification of the separation claim. revision: yes
Referee: [Abstract and Results] Abstract asserts empirical gains on curve fidelity, derivative preservation, downstream classification, and runtime across multiple benchmark datasets, yet supplies neither quantitative numbers, error bars, dataset descriptions, statistical tests, nor implementation details. This prevents any assessment of whether the claimed improvements are statistically meaningful or reproducible.

Authors: The abstract as written states the existence of gains without numerical values or statistical details, which limits immediate assessment. The full manuscript already contains the supporting tables with error bars, dataset descriptions, implementation details, and results on downstream tasks. To address the concern directly, we will revise the abstract to include representative quantitative improvements (e.g., relative reductions in reconstruction and derivative error) and will ensure that any statistical tests used in the experiments section are explicitly referenced. revision: yes

Circularity Check

0 steps flagged

No significant circularity; training-free method with no self-referential derivations or fitted predictions

full rationale

The paper describes a training-free Cascade-KDE procedure that estimates a 2D temporal-amplitude density directly from the input corrupted series, applies density truncation in a Robust Expectation step, and refines via exponential cascade. No equations, parameters, or results are presented that reduce by construction to the inputs (e.g., no fitted quantity renamed as a prediction, no self-citation chains justifying uniqueness, and no ansatz smuggled via prior work). Performance claims rest on empirical benchmarks rather than analytic derivations that loop back to the method's own definitions. This is the expected non-finding for a preprocessing heuristic without hidden self-dependencies.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

Abstract-only review yields no explicit free parameters or background axioms; the method is described as training-free. Two procedural components are treated as invented entities because they are introduced without external references.

invented entities (2)

Density-Truncated Robust Expectation no independent evidence
purpose: Limit influence of distant abnormal points during density-based restoration
Named as the second core stage of the pipeline; no prior reference supplied in abstract.
exponential cascade with adaptive stopping no independent evidence
purpose: Final refinement of the restored sequence
Named as the third stage; no external citation or derivation given.

pith-pipeline@v0.9.1-grok · 5716 in / 1196 out tokens · 55070 ms · 2026-06-30T16:00:46.212685+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 3 canonical work pages · 1 internal anchor

[1]

In: Proceedings of the AAAI conference on artificial intelligence

Chen, D., Lin, Y., Li, W., Li, P., Zhou, J., Sun, X.: Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 3438–3445 (2020)

2020
[2]

IEEE access9, 120043–120065 (2021)

Choi, K., Yi, J., Park, C., Yoon, S.: Deep learning for anomaly detection in time- series data: Review, analysis, and guidelines. IEEE access9, 120043–120065 (2021)

2021
[3]

arXiv preprint arXiv:2002.09545 (2020)

Gao, J., Song, X., Wen, Q., Wang, P., Sun, L., Xu, H.: Robusttad: Robust time series anomaly detection via decomposition and convolutional neural networks. arXiv preprint arXiv:2002.09545 (2020)

work page arXiv 2002
[4]

In: 2025 IEEE Applied Power Electronics Con- ference and Exposition (APEC)

Gong, T., Lee, J., Lee, S., Kim, Y., Ko, B., Na, W., Choi, S., Kim, J.: Enhanced incremental capacity analysis for evaluating battery degradation mechanisms of optimized fast charging methods. In: 2025 IEEE Applied Power Electronics Con- ference and Exposition (APEC). pp. 3006–3011. IEEE (2025)

2025
[5]

In: 2025 IEEE Vehicle Power and Propulsion Conference (VPPC)

Liao, B., Liu, Y., Gong, Y., Wu, Y., Sun, J.: State-of-charge estimation for battery based on integrated incremental capacity analysis and dual extended kalman filter. In: 2025 IEEE Vehicle Power and Propulsion Conference (VPPC). pp. 1–6. IEEE (2025)

2025
[6]

Frontiers of Information Technology & Electronic Engineering25(1), 19–41 (Dec 2023)

Lin, L., Li, Z., Li, R., Li, X., Gao, J.: Diffusion models for time-series ap- plications: a survey. Frontiers of Information Technology & Electronic Engineering25(1), 19–41 (Dec 2023). https://doi.org/10.1631/fitee.2300310, http://dx.doi.org/10.1631/FITEE.2300310

work page doi:10.1631/fitee.2300310 2023
[7]

In: Proceedings of the 33rd ACM Inter- national Conference on Information and Knowledge Management

Peng, J., Lei, R., Wei, Z.: Beyond over-smoothing: Uncovering the trainability challenges in deep graph neural networks. In: Proceedings of the 33rd ACM Inter- national Conference on Information and Knowledge Management. pp. 1878–1887 (2024)

2024
[8]

In: Meila, M., Zhang, T

Rasul, K., Seward, C., Schuster, I., Vollgraf, R.: Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8857–8868. PMLR (18–24 Jul 2021), https://proceedi...

2021
[9]

In: Proceed- ings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining

Su, Y., Zhao, Y., Niu, C., Liu, R., Sun, W., Pei, D.: Robust anomaly detection for multivariate time series through stochastic recurrent neural network. In: Proceed- ings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 2828–2837 (2019) 14 Y. Liu et al

2019
[10]

Advances in neural information processing systems34, 24804–24816 (2021)

Tashiro, Y., Song, J., Song, Y., Ermon, S.: Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in neural information processing systems34, 24804–24816 (2021)

2021
[11]

Ionics31(3), 2457–2471 (2025)

Wen, J., Wang, C., Zhao, Z., Sun, Z.: Capacity estimation of lithium-ion batter- ies based on segment ic curve data dimensionality reduction and reconstruction methods. Ionics31(3), 2457–2471 (2025)

2025
[12]

Journal of Electrochemical Energy Conversion and Storage22(4), 041012 (2025)

Yang, B., Huang, L., Wei, K., Shu, X., Garg, A., Li, Y.: A novel state of health estimation method for lithium-ion batteries based on incremental capacity curve. Journal of Electrochemical Energy Conversion and Storage22(4), 041012 (2025)

2025
[13]

In: Proceedings of the ACM Web Conference 2026

Zheng, Z., Yang, Y., Guan, Z., Zhao, W., Huang, X., Lu, W.: Beyond single- granularity prompts: A multi-scale chain-of-thought prompt learning for graph. In: Proceedings of the ACM Web Conference 2026. pp. 547–558 (2026)

2026
[14]

In: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2

Zheng, Z., Yang, Y., Guan, Z., Zhao, W., Lu, W.: Discrepancy-aware graph mask auto-encoder. In: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. pp. 4038–4049 (2025)

2025
[15]

In: Pro- ceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2

Zheng, Z., Yang, Y., Guan, Z., Zhao, W., Lu, W.: Enhancing homophily- heterophily separation: Relation-aware learning in heterogeneous graphs. In: Pro- ceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. pp. 4050–4061 (2025)

2025
[16]

Empowering Heterogeneous Graph Foundation Models via Decoupled Relation Alignment

Zheng, Z., Yang, Y., Wang, Z., Guan, Z., Zhao, W.: Empowering heteroge- neous graph foundation models via decoupled relation alignment. arXiv preprint arXiv:2605.00731 (2026), https://arxiv.org/abs/2605.00731

work page internal anchor Pith review Pith/arXiv arXiv 2026

[1] [1]

In: Proceedings of the AAAI conference on artificial intelligence

Chen, D., Lin, Y., Li, W., Li, P., Zhou, J., Sun, X.: Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 3438–3445 (2020)

2020

[2] [2]

IEEE access9, 120043–120065 (2021)

Choi, K., Yi, J., Park, C., Yoon, S.: Deep learning for anomaly detection in time- series data: Review, analysis, and guidelines. IEEE access9, 120043–120065 (2021)

2021

[3] [3]

arXiv preprint arXiv:2002.09545 (2020)

Gao, J., Song, X., Wen, Q., Wang, P., Sun, L., Xu, H.: Robusttad: Robust time series anomaly detection via decomposition and convolutional neural networks. arXiv preprint arXiv:2002.09545 (2020)

work page arXiv 2002

[4] [4]

In: 2025 IEEE Applied Power Electronics Con- ference and Exposition (APEC)

Gong, T., Lee, J., Lee, S., Kim, Y., Ko, B., Na, W., Choi, S., Kim, J.: Enhanced incremental capacity analysis for evaluating battery degradation mechanisms of optimized fast charging methods. In: 2025 IEEE Applied Power Electronics Con- ference and Exposition (APEC). pp. 3006–3011. IEEE (2025)

2025

[5] [5]

In: 2025 IEEE Vehicle Power and Propulsion Conference (VPPC)

Liao, B., Liu, Y., Gong, Y., Wu, Y., Sun, J.: State-of-charge estimation for battery based on integrated incremental capacity analysis and dual extended kalman filter. In: 2025 IEEE Vehicle Power and Propulsion Conference (VPPC). pp. 1–6. IEEE (2025)

2025

[6] [6]

Frontiers of Information Technology & Electronic Engineering25(1), 19–41 (Dec 2023)

Lin, L., Li, Z., Li, R., Li, X., Gao, J.: Diffusion models for time-series ap- plications: a survey. Frontiers of Information Technology & Electronic Engineering25(1), 19–41 (Dec 2023). https://doi.org/10.1631/fitee.2300310, http://dx.doi.org/10.1631/FITEE.2300310

work page doi:10.1631/fitee.2300310 2023

[7] [7]

In: Proceedings of the 33rd ACM Inter- national Conference on Information and Knowledge Management

Peng, J., Lei, R., Wei, Z.: Beyond over-smoothing: Uncovering the trainability challenges in deep graph neural networks. In: Proceedings of the 33rd ACM Inter- national Conference on Information and Knowledge Management. pp. 1878–1887 (2024)

2024

[8] [8]

In: Meila, M., Zhang, T

Rasul, K., Seward, C., Schuster, I., Vollgraf, R.: Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8857–8868. PMLR (18–24 Jul 2021), https://proceedi...

2021

[9] [9]

In: Proceed- ings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining

Su, Y., Zhao, Y., Niu, C., Liu, R., Sun, W., Pei, D.: Robust anomaly detection for multivariate time series through stochastic recurrent neural network. In: Proceed- ings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 2828–2837 (2019) 14 Y. Liu et al

2019

[10] [10]

Advances in neural information processing systems34, 24804–24816 (2021)

Tashiro, Y., Song, J., Song, Y., Ermon, S.: Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in neural information processing systems34, 24804–24816 (2021)

2021

[11] [11]

Ionics31(3), 2457–2471 (2025)

Wen, J., Wang, C., Zhao, Z., Sun, Z.: Capacity estimation of lithium-ion batter- ies based on segment ic curve data dimensionality reduction and reconstruction methods. Ionics31(3), 2457–2471 (2025)

2025

[12] [12]

Journal of Electrochemical Energy Conversion and Storage22(4), 041012 (2025)

Yang, B., Huang, L., Wei, K., Shu, X., Garg, A., Li, Y.: A novel state of health estimation method for lithium-ion batteries based on incremental capacity curve. Journal of Electrochemical Energy Conversion and Storage22(4), 041012 (2025)

2025

[13] [13]

In: Proceedings of the ACM Web Conference 2026

Zheng, Z., Yang, Y., Guan, Z., Zhao, W., Huang, X., Lu, W.: Beyond single- granularity prompts: A multi-scale chain-of-thought prompt learning for graph. In: Proceedings of the ACM Web Conference 2026. pp. 547–558 (2026)

2026

[14] [14]

In: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2

Zheng, Z., Yang, Y., Guan, Z., Zhao, W., Lu, W.: Discrepancy-aware graph mask auto-encoder. In: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. pp. 4038–4049 (2025)

2025

[15] [15]

In: Pro- ceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2

Zheng, Z., Yang, Y., Guan, Z., Zhao, W., Lu, W.: Enhancing homophily- heterophily separation: Relation-aware learning in heterogeneous graphs. In: Pro- ceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. pp. 4050–4061 (2025)

2025

[16] [16]

Empowering Heterogeneous Graph Foundation Models via Decoupled Relation Alignment

Zheng, Z., Yang, Y., Wang, Z., Guan, Z., Zhao, W.: Empowering heteroge- neous graph foundation models via decoupled relation alignment. arXiv preprint arXiv:2605.00731 (2026), https://arxiv.org/abs/2605.00731

work page internal anchor Pith review Pith/arXiv arXiv 2026