Lag-aware cross-hand alignment for dual-hand action segmentation

Fatemeh Ziaeetabar

REVIEW 1 major objections 6 minor 16 references

Lag-aware cross-hand alignment for dual-hand action segmentation

T0 review · 1 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read LACA shows that explicitly aligning time-varying offsets between hands improves dual-hand action segmentation and boundary localization.

desk verdict A genuinely new but incremental alignment module with a careful evaluation; the main risk is the sensitivity of its hand-tuned target construction, which deserves a sensitivity analysis before publication. read the letter →

arxiv 2607.26215 v1 pith:GGMBNNHA submitted 2026-07-28 cs.CV

Fatemeh Ziaeetabar This is my paper

classification cs.CV

keywords dual-handactionsegmentationtemporalalignmentcross-handcoordinationlagestimationboundarylocalizationfuture-freebimanualinteractiontransitionsupervision

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

The paper argues that dual-hand action segmentation fails when it fuses left- and right-hand features at the same time step, because coordinated hand transitions often occur with nonzero, time-varying delays. It introduces Lag-Aware Cross-Hand Alignment (LACA), a module that learns a directional probability distribution over possible temporal offsets between the two hands' feature streams and retrieves the opposite-hand feature from the estimated offset. A learned null state suppresses cross-hand transfer when no compatible transition exists, and training targets are mined automatically from existing frame-level annotations. On two assembly-video benchmarks, adding LACA to a strong baseline improves two-hand F1@50 by about two points and boundary F1 by about three points, with negligible added parameters. A future-free variant, LACA-C, reduces the median delay of transition cues from 300 ms to 233 ms while raising transition recall.

What carries the argument

The central object is LACA, a module that computes, for each hand and time step, a compatibility score between the target hand's query and candidate opposite-hand features at offsets −K..K, jointly with a learned null key. These scores are softmax-normalized to form a directional lag distribution; the aligned cross-hand representation is a probability-weighted sum of retrieved features, and the null probability gates how much of that representation is added to the target stream. Supervision comes from compatibility-aware targets: transition types are matched across hands using co-occurrence statistics from training annotations, and only accepted matches receive Gaussian soft lag targets; unm

What would settle it

Measure the true cross-hand transition lag distribution on a held-out bimanual dataset; if the median absolute lag exceeds K sampled steps (e.g., >15 steps at 7.5 Hz, i.e., >2 seconds), then LACA's improvement over same-index fusion would vanish or reverse, since its search window cannot cover the relevant offset.

Watch

Extended reading notes

Core claim

The central claim is that explicit, learned cross-hand temporal alignment—rather than larger temporal windows or implicit fusion—is what carries the improvement. The paper demonstrates this by ablating its module: local temporal attention searching the same offset window without lag supervision or null state gains only 0.6 F1@50 points, while full LACA gains 2.1 points. The paper also reports that 44.7% (HA-ViD) and 48.9% (ATTACH) of transition anchors have robust nonzero cross-hand matches, versus 18.6% and 21.3% under temporally shifted controls, indicating real, structured inter-hand delay.

Load-bearing premise

The matching rule assumes that related cross-hand transitions lie within K=15 sampled steps of each other and that compatibility statistics computed from training annotations transfer to test sequences; if real lags exceed K, or the hand-tuned thresholds reject valid matches, the lag targets become noisy or empty and the benefit disappears.

Editorial extensions

If this is right

Adding LACA to a diffusion-based dual-hand baseline improves two-hand mean F1@50 from 40.4 to 42.5 on HA-ViD and from 19.9 to 21.8 on ATTACH, and boundary F1 from 56.5 to 59.6 and 44.7 to 47.9 respectively, with only about 0.0086M extra parameters.
Ablations show that both lag supervision and the null state contribute: removing lag supervision drops F1@50 by 0.9 points; removing the null state drops B-F1 by 1.9 points.
A future-free variant, LACA-C, achieves 83.6% transition-cue recall, 233 ms median availability delay, and 0.72 false cues per minute on ATTACH, compared with 76.4%, 300 ms, and 1.15 for the future-free baseline.
Annotation analysis shows robust nonzero cross-hand matches at 44.7% and 48.9% of transition anchors versus 18.6% and 21.3% under shifted controls, suggesting structured cross-hand lag.

Reading between the lines

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

If real-world bimanual actions have lags exceeding the fixed K=15-step window, LACA's benefit will degrade; a testable extension is to make K adaptive or to estimate a global lag prior per action type.
The null-state mechanism could be repurposed as a confidence signal for when cross-hand context is unlikely to matter, potentially informing when to skip cross-hand computation for efficiency.
The compatibility-aware target construction from co-occurrence statistics could be applied to other multi-stream alignment tasks (e.g., audio-visual event alignment) without additional labels.
Since LACA-C still trails offline LACA by 0.7 F1@50, there is room for future-free variants that recover part of that gap, for instance by predictive context rather than strict nonpositive offsets.

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.

Desk Editor's Note

read the letter

Dear colleague,

You should know that this is a solid, honest paper about a small but real improvement. It introduces LACA, a module that explicitly estimates directional, time-varying cross-hand offsets with a learned null state, instead of fusing left/right features at the same time index as Polyphony and other baselines do. That is genuinely new — I am not aware of prior work that estimates per-frame lag distributions with a null state for dual-hand segmentation. The compatibility-aware target construction from training annotations is also clever and does not require additional labels.

The paper is careful. Three seeds, validation-based checkpoint selection, reserved test set, and a permutation control to show the annotation lag statistics are not an artifact of the matching rule. The ablation isolates the lag supervision and the null state. Gains are modest: about 2 points F1@50 and 3 points B-F1 on both datasets, with negligible parameter cost. The boundary F1 improvement is consistent with the motivation.

The soft spots are real but not disqualifying. The target construction in Eq. (12) depends on hand-tuned thresholds θ=0.20, α=0.30, and window K=15. As the stress-test note says, a small change in θ could flip many anchors between a Gaussian target and the null state. The paper does not report the fraction of anchors assigned to the null state, nor a sensitivity sweep. The ablation shows lag supervision contributes only about 0.9 F1@50 over the null-gated model, so if a plausible parameter setting erodes that margin, the central claim weakens. The authors acknowledge the fixed search range and limited datasets in the Conclusion, which is honest, and the future-free variant is tested only on ATTACH. Code is not released, which makes it harder to probe these questions.

None of this, in my view, sinks the paper. The evaluation is careful enough that the reader's conditional verdict — accept if sensitivity is addressed and code is shared — is the right one. The paper is worth a serious referee.

I would bring it to a reading group as a good example of a focused, well-evaluated incremental technique, and I would cite it if I worked in temporal action segmentation.

Recommendation: send it to peer review, with a request for sensitivity analysis on the target-construction hyperparameters and release of code.

Referee Report

1 major / 6 minor

Summary. The paper introduces LACA, a lightweight module for dual-hand action segmentation that replaces same-index cross-hand fusion in Polyphony with attention over a bounded set of temporal offsets in both directions. A learned null key gates cross-hand retrieval, and lag supervision is derived automatically from frame-level training labels via a compatibility-aware matching rule (Eq. 12) that accepts only temporally proximal transition pairs with sufficiently high smoothed transition-type compatibility. The authors analyze transition-lag statistics on HA-ViD and ATTACH, report F1@50 and boundary-F1 gains over a reproduced Polyphony baseline, and present a future-free variant (LACA-C) evaluated on ATTACH with right-aligned features and nonpositive offsets.

Significance. If the reported gains hold, LACA addresses a real and under-modeled issue—asynchronous bimanual transitions—with negligible added parameters and no extra annotation. The evaluation is diligent: three seeds with mean±SD, validation-based checkpoint/hyperparameter selection with the test set reserved, a controlled reproduction of Polyphony, ablations separating the search window, lag supervision, and null state, a permutation-based annotation control, and a carefully constructed future-free protocol. However, the central mechanism's targets depend on hand-set constants, and the missing sensitivity analysis leaves open the possibility that the ~2-point gains are an artifact of those settings rather than of explicit alignment. The contribution is therefore promising but not yet fully supported; the required analysis is within the scope of a revision.

major comments (1)

[Section 3.3, Eq. (12); Section 4.1; Section 4.4, Table 2] LACA's supervision depends on the hard acceptance rule with freely chosen K=15, alpha=0.30, theta=0.20, epsilon=1.0, yet no sensitivity analysis or accepted-anchor/null fraction is reported. With C smoothed over many transition types, most compatibility scores are near 1/|R|, so theta controls whether anchors receive a Gaussian soft target or are assigned to null. Table 2's ablation credits lag supervision with only ~0.9 F1@50 and ~1.7 B-F1 over LACA w/o Llag, and local temporal attention with only 0.6 F1@50; a modest shift in theta or K could plausibly erase this margin. Please report acceptance/null fractions and a validation sweep over theta and K, and preferably alpha and sigma, showing both downstream F1@50/B-F1 and the RRNZ rates.

minor comments (6)

[Section 4.1] eta=0.50 is listed among the hyperparameters but never defined in the text; if it is the adaptive hand-weighting coefficient, state this and give the formula.
[Section 4.2] With N_perm=1000, a permutation p-value or a 95% CI for the shifted-control rates would be a one-line addition; as written, 'observed vs mean of controls' lacks a measure of uncertainty. The RRNZ rate is also conditional on K=15 and rho=2, so a K-sweep would strengthen the descriptive claim.
[Sections 3.5/4.5 and Abstract] Future-free evaluation is conducted only on ATTACH; the abstract's 'supporting timely future-free perception' is a broader claim. The Conclusion acknowledges this, but the abstract and contribution list should state the single-dataset scope.
[Abstract; Section 3] LACA is described as a 'backbone-independent module', but only the Polyphony instantiation is evaluated. Please qualify this claim or add a second backbone experiment.
[Table 2; Section 4.4] The active-parameter increase from Polyphony (18.760 M) to LACA (18.769 M) is 0.009 M at the displayed precision, while the text says 0.0086 M. Reconcile by reporting more decimals or rounding consistently.
[Overall] There is no code-release statement. Given the sensitivity of the target construction to theta and K, releasing the target-generation code and the exact hyperparameter selection protocol would greatly aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: LACA's lag targets come from training annotations and its gains are measured on held-out test data against a reproduced baseline.

full rationale

The paper's central claim is that explicit cross-hand temporal alignment improves dual-hand action segmentation. The lag-supervision targets in Eq. (12)-(13) are constructed from ground-truth frame-level training labels via a compatibility heuristic (Eq. (10)-(11)) that is independent of the model's own outputs, and the paper states that 'Test annotations are never used for compatibility estimation, target construction, checkpoint selection, or hyperparameter tuning.' The reported improvements (F1@50 40.4 to 42.5, B-F1 56.5 to 59.6 on HA-ViD; 19.9 to 21.8 and 44.7 to 47.9 on ATTACH) compare LACA against a reproduced Polyphony baseline on test splits, so the central result is not defined in terms of its own inputs. The ablation shows the null state and lag supervision are each load-bearing relative to a same-window local attention baseline, and the empirical lag-structure analysis is descriptive rather than a fitted prediction. The only self-citations ([7], [8], [9]) appear in related work and are not used to justify the proposed method or to forbid alternatives. The skeptic's concern about hand-tuned thresholds (alpha, theta, K) and missing sensitivity analysis is a legitimate robustness/correctness concern, not a circularity: the paper does not redefine the test metric in terms of those thresholds, and it does not claim the thresholds were derived from the evaluated predictions. No step in the derivation reduces, by construction or by self-citation, to its own input.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several hand-chosen hyperparameters (K, α, θ, σ, λ_ℓ) and on the assumption that the compatibility-aware matching rule captures genuine cross-hand relationships. The method introduces no new physical or conceptual entities beyond the learned null key and offset distributions, which are standard components of attention-based models.

free parameters (8)

K (search window half-size) = 15 sampled steps (~2 s at 7.5 Hz)
Chosen by hand; defines the set of candidate offsets for alignment. Performance depends on this range.
α (distance penalty coefficient) = 0.30
Chosen by hand; weights temporal distance in the matching score of Eq. (12).
θ (match acceptance threshold) = 0.20
Chosen by hand; determines whether a candidate match is accepted or the null state is used.
σ (Gaussian target width) = 2.0 sampled steps
Chosen by hand; controls the softness of the lag supervision targets in Eq. (13).
λ_ℓ (lag loss weight) = 0.20
Chosen by hand; balances the segmentation objective and lag supervision in Eq. (14).
η (adaptive hand weighting or similar) = 0.50
Mentioned in the hyperparameter list but not fully described in the text; likely part of the Polyphony objective.
ρ (robust-nonzero threshold) = 2 sampled steps
Chosen by hand; defines what counts as a 'robust nonzero' lag in the annotation analysis of Eq. (18).
ϵ (compatibility smoothing) = 1.0
Chosen by hand; additive smoothing in the compatibility estimate of Eq. (11).

assumptions (5)

standard math Scaled dot-product attention (Vaswani et al.) is a valid mechanism for comparing and retrieving features.
The alignment module is built on attention; the paper relies on its well-known properties without proof.
domain assumption Related cross-hand transitions lie within a bounded temporal window of ±K sampled steps.
The search space for alignment is restricted to Δ_K; if real lags exceed K, the method cannot retrieve them.
domain assumption The temporal offset between related transitions can be approximated by a single Gaussian-peaked target centered at the best match δ*.
Eq. (13) constructs a unimodal Gaussian target, assuming the offset distribution is not multi-modal or heavy-tailed.
domain assumption Transition-type compatibility estimated from training annotations generalizes to test videos.
The matching rule in Eq. (12) uses compatibility matrices C from D_train; if these are not representative, targets become misleading.
domain assumption For future-free inference, nonpositive offsets (current or past) suffice to provide useful cross-hand context.
LACA-C restricts offsets to Δ_C; this assumes no benefit from looking ahead, which may not hold if the opposite hand leads.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Lag-aware cross-hand alignment for dual-hand action segmentation." pith.science (2026). https://pith.science/paper/GGMBNNHA

@misc{pith2026260726215,
  author       = {Pith},
  title        = {Pith review of: Lag-aware cross-hand alignment for dual-hand action segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GGMBNNHA}},
  note         = {Machine review of arXiv:2607.26215}
}

read the original abstract

Dual-hand action segmentation commonly fuses left- and right-hand representations at identical temporal indices, although coordinated hand transitions may occur with nonzero and time-varying delays. We introduce Lag-Aware Cross-Hand Alignment (LACA), a lightweight module that explicitly estimates directional temporal-offset distributions between hand-specific feature streams. LACA retrieves cross-hand information from the estimated offsets and incorporates a learned null state to suppress transfer when no compatible cross-hand transition is supported. Alignment is supervised using compatibility-aware targets derived automatically from frame-level training annotations, without requiring additional labels. Analysis of the HA-ViD and ATTACH training annotations reveals robust nonzero cross-hand matches for 44.7% and 48.9% of transition anchors, respectively, compared with 18.6% and 21.3% under temporally shifted controls. When integrated into Polyphony, LACA improves the two-hand mean F1@50 from 40.4 to 42.5 and boundary F1 from 56.5 to 59.6 on HA-ViD, and from 19.9 to 21.8 and 44.7 to 47.9, respectively, on ATTACH, relative to our reproduced Polyphony baseline. These gains require only approximately 0.0086 million additional trainable parameters. We further introduce LACA-C, a future-free variant that restricts alignment and the complete inference pipeline to current and past observations. On ATTACH, LACA-C achieves 83.6% transition-cue recall, a seed-averaged median availability delay of 233~ms, 0.72 false cues per minute, and segmentation-stage throughput of 224.9 current-position predictions per second. These results demonstrate that explicit cross-hand temporal alignment improves both action segmentation and boundary localization while supporting timely future-free perception.

Figures

Figures reproduced from arXiv: 2607.26215 by the authors.

**Figure 1.** Motivation and overview of lag-aware cross-hand alignment. (a) Left- and right-hand label timelines in an assembly sequence, where the left-hand [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Empirical lag validation and future-free evaluation. (a) Observed robust-nonzero transition-anchor rates on HA-ViD and ATTACH compared with the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 1 canonical work pages

[1]

Matheson, R

E. Matheson, R. Minto, E. G. G. Zampieri, M. Faccio, G. Rosati, Human–robot collaboration in manufacturing applications: A review, Robotics 8 (4) (2019) 100.doi: 10.3390/robotics8040100

doi:10.3390/robotics8040100 2019
[2]

Aganian, B

D. Aganian, B. Stephan, M. Eisenbach, C. Stretz, H.- M. Gross, ATTACH dataset: Annotated two-handed assembly actions for human action understanding, in: 2023 IEEE International Conference on Robotics and Automation, 2023, pp. 11367–11373.doi:10.1109/ ICRA48891.2023.10160633

arXiv 2023
[3]

Zheng, H

H. Zheng, H. Wang, T. Zheng, P. Bhattarai, T. Alhanai, Polyphony: Diffusion-based dual-hand action segmenta- tion with alternating vision transformer and semantic con- ditioning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 20098–20107

2026
[4]

Zheng, R

H. Zheng, R. Lee, Y . Lu, HA-ViD: A human assembly video dataset for comprehensive assembly knowledge un- derstanding, in: Advances in Neural Information Process- ing Systems, V ol. 36, 2023, pp. 67069–67081

2023
[5]

Zheng, R

H. Zheng, R. Lee, Y . Lu, X. Xu, DuHa: A dual-hand ac- tion segmentation method for human–robot collaborative assembly, in: 2024 IEEE 20th International Conference on Automation Science and Engineering, 2024, pp. 522– 527.doi:10.1109/CASE59546.2024.10711776

arXiv 2024
[6]

Zheng, R

H. Zheng, R. Lee, H. Liang, Y . Lu, X. Xu, DuCAS: A knowledge-enhanced dual-hand compositional action segmentation method for human–robot collaborative as- sembly, in: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024, pp. 7175–7180. doi:10.1109/IROS58592.2024.10802758

arXiv 2024
[7]

Ziaeetabar, M

F. Ziaeetabar, M. Tamosiunaite, F. Wörgötter, A hierarchi- cal graph-based approach for recognition and description generation of bimanual actions in videos, IEEE Access 12 (2024) 180328–180360.doi:10.1109/ACCESS.2024. 3509674

doi:10.1109/access.2024 2024
[8]

Ziaeetabar, F

F. Ziaeetabar, F. Wörgötter, Adaptive multimodal graph reasoning with foundation models for fine-grained ac- tion recognition, IEEE Access 13 (2025) 201990–202009. doi:10.1109/ACCESS.2025.3637990

arXiv 2025

Show all 16 references

[9]

Ziaeetabar, Efficientgformer: Multimodal brain tumor segmentation via pruned graph-augmented transformer, arXiv preprint arXiv:2508.01465 (2025)

F. Ziaeetabar, Efficientgformer: Multimodal brain tumor segmentation via pruned graph-augmented transformer, arXiv preprint arXiv:2508.01465 (2025)

2025 arXiv
[10]

Zhong, G

Q. Zhong, G. Ding, A. Yao, OnlineTAS: An online base- line for temporal action segmentation, in: Advances in Neural Information Processing Systems, V ol. 37, 2024. doi:10.52202/079017-1881

2024 doi
[11]

P. Lei, S. Todorovic, Temporal deformable residual net- works for action segmentation in videos, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 6742–6751

2018
[12]

Chang, D.-A

C.-Y . Chang, D.-A. Huang, Y . Sui, L. Fei-Fei, J. C. Niebles, D 3TW: Discriminative differentiable dynamic 7 time warping for weakly supervised action alignment and segmentation, in: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2019, pp...

2019
[13]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need, in: Advances in Neural Information Processing Systems, V ol. 30, 2017, pp. 5998–6008

2017
[14]

Abu Farha, J

Y . Abu Farha, J. Gall, MS-TCN: Multi-stage temporal convolutional network for action segmentation, in: Pro- ceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2019, pp. 3575–3584.doi: 10.1109/CVPR.2019.00369

2019
[15]

D. Liu, Q. Li, A.-D. Dinh, T. Jiang, M. Shah, C. Xu, Diffusion action segmentation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 10139–10149

2023
[16]

Z. Lu, E. Elhamifar, FACT: Frame-action cross-attention temporal modeling for efficient action segmentation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18175–18185. 8

2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Matheson, R

E. Matheson, R. Minto, E. G. G. Zampieri, M. Faccio, G. Rosati, Human–robot collaboration in manufacturing applications: A review, Robotics 8 (4) (2019) 100.doi: 10.3390/robotics8040100

doi:10.3390/robotics8040100 2019

[2] [2]

Aganian, B

D. Aganian, B. Stephan, M. Eisenbach, C. Stretz, H.- M. Gross, ATTACH dataset: Annotated two-handed assembly actions for human action understanding, in: 2023 IEEE International Conference on Robotics and Automation, 2023, pp. 11367–11373.doi:10.1109/ ICRA48891.2023.10160633

arXiv 2023

[3] [3]

Zheng, H

H. Zheng, H. Wang, T. Zheng, P. Bhattarai, T. Alhanai, Polyphony: Diffusion-based dual-hand action segmenta- tion with alternating vision transformer and semantic con- ditioning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 20098–20107

2026

[4] [4]

Zheng, R

H. Zheng, R. Lee, Y . Lu, HA-ViD: A human assembly video dataset for comprehensive assembly knowledge un- derstanding, in: Advances in Neural Information Process- ing Systems, V ol. 36, 2023, pp. 67069–67081

2023

[5] [5]

Zheng, R

H. Zheng, R. Lee, Y . Lu, X. Xu, DuHa: A dual-hand ac- tion segmentation method for human–robot collaborative assembly, in: 2024 IEEE 20th International Conference on Automation Science and Engineering, 2024, pp. 522– 527.doi:10.1109/CASE59546.2024.10711776

arXiv 2024

[6] [6]

Zheng, R

H. Zheng, R. Lee, H. Liang, Y . Lu, X. Xu, DuCAS: A knowledge-enhanced dual-hand compositional action segmentation method for human–robot collaborative as- sembly, in: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024, pp. 7175–7180. doi:10.1109/IROS58592.2024.10802758

arXiv 2024

[7] [7]

Ziaeetabar, M

F. Ziaeetabar, M. Tamosiunaite, F. Wörgötter, A hierarchi- cal graph-based approach for recognition and description generation of bimanual actions in videos, IEEE Access 12 (2024) 180328–180360.doi:10.1109/ACCESS.2024. 3509674

doi:10.1109/access.2024 2024

[8] [8]

Ziaeetabar, F

F. Ziaeetabar, F. Wörgötter, Adaptive multimodal graph reasoning with foundation models for fine-grained ac- tion recognition, IEEE Access 13 (2025) 201990–202009. doi:10.1109/ACCESS.2025.3637990

arXiv 2025