pith. sign in

arxiv: 2605.22597 · v1 · pith:7EF6EAS2new · submitted 2026-05-21 · 💻 cs.LG · cs.AI· cs.GR· cs.RO

MoSA: Motion-constrained Stress Adaptation for Mitigating Real-to-Sim Gap in Continuum Dynamics via Learning Residual Anisotropy

Jiaxu Wang , Junhao He , Jingkai Sun , Yi Gu , Yunyang Mo , Jiahang Cao , Qiang Zhang , Renjing Xu This is my paper

Pith reviewed 2026-05-22 06:54 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.GRcs.RO
keywords real-to-sim gapcontinuum dynamicsresidual anisotropyphysics-informed learningstress adaptationdeformable objectsrobot manipulationmicroplane redistribution
0
0 comments X

The pith

MoSA learns residual stress operators on an isotropic backbone to capture mild anisotropy and close the real-to-sim gap in continuum dynamics.

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

The paper targets the limit that arises once an isotropic simulator has been calibrated to real objects: real materials still show mild anisotropy and heterogeneity that the base model cannot represent. MoSA keeps the calibrated isotropic physics as a strong prior and adds learnable residual stress operators inside a cascaded network. These operators adapt the stress field step by step through microplane-constrained redistribution while motion constraints supervise the temporal and spatial derivatives of the deformation. The resulting hybrid model produces more accurate and generalizable dynamics from visual data than either pure physics or black-box neural approaches. The same improvement carries over to more reliable sim-to-real transfer when the learned dynamics are used for robot manipulation planning.

Core claim

After the near-isotropic backbone is well calibrated, the remaining real-to-sim gap is dominated by mild residual anisotropy and heterogeneity. MoSA captures these effects by learning residual stress operators that progressively adapt stresses via microplane-constrained redistribution inside a physics-informed cascaded network, while motion constraints are enforced by supervising derivatives of the deformation field. The result is higher accuracy, better generalization, and physically interpretable residual anisotropy that improves downstream sim-to-real robot performance.

What carries the argument

Residual stress operators that perform microplane-constrained redistribution of stresses inside a cascaded network, built on top of a calibrated isotropic model and regularized by motion constraints on the deformation field.

If this is right

  • Dynamics learned this way achieve higher accuracy and stronger generalization than either the isotropic baseline or end-to-end neural models.
  • The learned residual fields remain physically meaningful and can be inspected for anisotropy patterns.
  • Improved real-to-sim dynamics directly increase the reliability of sim-to-real transfer for robot manipulation tasks.
  • The method retains data efficiency and physical consistency by never discarding the isotropic prior.

Where Pith is reading between the lines

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

  • The same residual-operator pattern could be tried in other simulation domains where the base model is already close but needs small structured corrections.
  • Structured residual learning may be especially useful when the expected discrepancy is mild and physically interpretable rather than arbitrary.
  • Testing the framework on materials with stronger heterogeneity or large-deformation regimes would reveal where the mild-anisotropy assumption stops holding.

Load-bearing premise

Once the isotropic model is calibrated, the leftover real-to-sim error is mainly mild residual anisotropy and heterogeneity that can be represented by learnable stress operators without overfitting or breaking physical consistency.

What would settle it

If adding the learned residual stress operators produces no reduction in prediction error on held-out real deformation sequences, or if the adapted stresses cause unstable or non-physical simulations, the claim that these operators close the gap would be falsified.

Figures

Figures reproduced from arXiv: 2605.22597 by Jiahang Cao, Jiaxu Wang, Jingkai Sun, Junhao He, Qiang Zhang, Renjing Xu, Yi Gu, Yunyang Mo.

Figure 1
Figure 1. Figure 1: Overview of the pipeline. (a) Two-stage dynamic reconstruction. (b)Simulation with progressive anisotropic stress adaptation (c) Motion-constrained optimization strategy 3. Methodology Problem Definition. This work aims to learn real-to-sim dynamics from multi-view videos by preserving an isotropic constitutive prior as the backbone and modeling the resid￾ual effects induced by mild anisotropy and heteroge… view at source ↗
Figure 2
Figure 2. Figure 2: Rescaling and redistributing of stress tensor complementary matrix since we assume that the prior con￾stitutive equation, even when describing isotropy, still has some deficiencies. Therefore, it is expected to make slight adjustments to the isotropic stress. We produce this term by: L pre = WpreΦ(U, σp) + b pre , (4) where Φ is an MLP with two hidden layers. σp refers to the prior stress σp = σ(ϵ). U is t… view at source ↗
Figure 4
Figure 4. Figure 4: Comparisons of long-term dynamics (a) and sensitivity of physical laws (b). 4.2. Real-world Generalization Real-world evaluation is essential because, even when an isotropic backbone is a reasonable approximation, common objects still exhibit mild anisotropy and heterogeneity that can dominate the remaining real-to-sim error. In this ex￾periment, we learn dynamics from video recordings and evaluate general… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative Comparisons between our methods and baselines. More Visualization can be seen in the Appendix. Distance (CD) and Earth Mover Distance (EMD) for 3D geometry error, and PSNR and SSIM for rendering quality. For PSNR and SSIM, we crop images around the object to avoid background dominance. On the real-world dataset, we report only PSNR and SSIM since ground-truth 3D particle trajectories are unavai… view at source ↗
Figure 5
Figure 5. Figure 5: Applications of zero-shot robot manipulation transfer. parameters to compensate for model errors. In contrast, our method refines the constitutive prior itself, making it less dependent on model selection. GIC exhibits large perfor￾mance variation across different priors, whereas our method remains stable. Notably, even starting from a simple linear stress–strain model, our approach can progressively corre… view at source ↗
Figure 7
Figure 7. Figure 7: Spatial visualization of the normalized learned hetero￾geneity field η(x) on test objects. Reddish regions correspond to larger learned local stiffness, while bluish regions indicate softer regions. visualize the normalized learned η(x) on three held-out test objects in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Physically grounded analysis of learned anisotropy A. Verification of Learned Anisotropy Knowledge To verify that the proposed physics-informed Progressive Stress Adaptation module truly learns the underlying physical anisotropy rather than merely benefiting from increased parameters, we design two validation experiments based on the same simulation. A uniaxial compression test is conducted on a cylindrica… view at source ↗
Figure 9
Figure 9. Figure 9: Grid Search for the hyperparameter α1,2 E. Effect of the Prior Stress We evaluate the effect of removing prior stress, specifically σ(ϵ)kl in Eq. 3 of the main paper, and relying only on C and L (as shown in [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual illustration of our real-world data collection pipeline (a) Chick (b) Rabbit (c) Mandarin (d) Chick2 (e) Peanut (f) Gorilla (g) RainbowBall [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The visualizations of all objects in our real-world dataset To achieve high-quality, real-world dynamic data, we utilize an advanced light field reconstruction system, as shown in Fig.10, to gather multi-view RGB sequences. This system is equipped with ultra-high-precision industrial cameras arranged in a spherical configuration, allowing for synchronized, surround-view data collection with a delay of les… view at source ↗
Figure 12
Figure 12. Figure 12: Comparisons of simulation results on PAC-NeRF dataset with or without metric regularization (e.g. Lscale, Lrot) J. Explanation of Scale Regularization In this section, we establish the relationship between the scale changes of Gaussian splats and the simulated deformation gradient. Since the static Gaussians are assumed to be isotropic, all anisotropic deformations during dynamic reconstruction are captur… view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparison of different baselines and our methods on the chick2 scenario. Ours GT GIC NeuMA DEL Vid2Sim [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison of different baselines and our methods on the chick1 scenario. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison of different baselines and our methods on the rabbit scenario. Ours GT GIC NeuMA DEL Vid2Sim [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative comparison of different baselines and our methods on the peanut scenario. O. Derivation of Microplane Parametrization of the Fourth-Order Correction Operator In this appendix, we show, in a self-contained way, that the proposed microplane parametrization using (Cx, Cy, Cz, Cxyz) is mathematically equivalent to using a single fourth-order tensor Cijkl acting on the stress tensor in Eq. (3). The… view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative comparison of different baselines and our methods on the gorilla scenario. Fourth-order tensor contraction in Voigt form. The core correction in Eq. 3 can be written as σˆij = σij + Cijkl σkl, (19) where Cijkl is a fourth-order tensor. Because σij is symmetric, we can collect the six independent components into a Voigt vector σ =         σxx σyy σzz σyz σzx σxy         , σˆ = … view at source ↗
Figure 18
Figure 18. Figure 18: Qualitative comparison of different baselines and our methods on the Rainbowball scenario. which is exactly what Eq. 5 does. Each entry of σ˜ x , for example, is just a linear combination of (σxx, σxy, σxz). Stacking all intermediate components back into a Voigt vector σ (1) (and using only index reordering), this whole “within-plane” redistribution is again a 6 × 6 matrix acting on σ: σ (1) = Bplane σ, B… view at source ↗
read the original abstract

Learning real-world dynamics from visual observations is crucial for various domains. A common strategy is to calibrate simulators by estimating physical parameters, yet accuracy is ultimately bounded by the underlying physical models, which often assume materials are homogeneous and isotropic. Even if reasonable, real-world objects typically exhibit mild anisotropy and heterogeneity. After the near-isotropic backbone is well calibrated, these residual effects become the key bottleneck for further closing the real-to-sim gap. Although neural networks can fit dynamics end-to-end, such black-box modeling discards strong physical priors, leading to poor data efficiency and overfitting. Therefore, we propose MoSA, a motion-constrained stress adaptation framework that targets these residual effects to further improve real-to-sim dynamics learning. MoSA uses an isotropic model as a physics prior and learns residual stress operators to capture mild anisotropy and heterogeneity. It progressively adapts stresses via microplane-constrained redistribution in a physics-informed cascaded network. We further impose motion constraints by supervising temporal and spatial derivatives of the deformation field. Experimentally, our learned dynamics achieves superior accuracy, generalization, and robustness, while learning physically meaningful residual anisotropy. Finally, we validate MoSA in a robot manipulation setting, showing that better real-to-sim dynamics modeling translates into more reliable sim-to-real transfer. Project Page is available at https://mercerai.github.io/MoSA/.

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

3 major / 3 minor

Summary. The manuscript proposes MoSA, a motion-constrained stress adaptation framework for learning continuum dynamics from visual observations. It uses an isotropic model as a physics prior and learns residual stress operators to capture mild anisotropy and heterogeneity via microplane-constrained redistribution in a physics-informed cascaded network, with additional supervision on temporal and spatial derivatives of the deformation field to enforce motion constraints. The central claims are that this yields superior accuracy, generalization, and robustness while producing physically meaningful residuals, and that the improved dynamics modeling enhances sim-to-real transfer in robot manipulation tasks.

Significance. If the empirical claims hold, the work provides a structured way to close residual real-to-sim gaps by augmenting calibrated isotropic priors with learned, constrained residuals rather than discarding physics entirely. This could improve data efficiency and physical consistency in learned dynamics models for deformable objects, with clear relevance to robotics and simulation-based control. The explicit use of motion constraints and microplane redistribution to preserve consistency is a positive design choice that merits further exploration if supported by results.

major comments (3)
  1. [Abstract] Abstract: the claims of 'superior accuracy, generalization, and robustness' and 'physically meaningful residual anisotropy' are presented without any quantitative error metrics, ablation results, baseline comparisons, or validation statistics. This absence is load-bearing for the central empirical claim that residual stress operators close the remaining real-to-sim gap after isotropic calibration.
  2. [Method] Method section on residual formulation: the weakest assumption—that mild residual anisotropy and heterogeneity dominate the post-calibration gap and can be captured without overfitting or violating consistency—requires explicit support. No analysis is visible showing that other sources (e.g., contact modeling, sensor noise, or discretization) are secondary, nor are there controls demonstrating that the learned operators remain physically admissible under the imposed motion constraints.
  3. [Experiments] Robot manipulation validation: the statement that better dynamics modeling 'translates into more reliable sim-to-real transfer' needs quantitative transfer metrics (success rates, trajectory error, or failure modes) with and without MoSA to establish the practical impact; qualitative demonstration alone is insufficient for the claim.
minor comments (3)
  1. [Method] Clarify the exact parameterization of the residual stress operator (e.g., its functional form, number of learnable weights, and how microplane constraints are enforced at each cascade step) to support reproducibility.
  2. [Related Work] Add a short discussion or reference to related work on microplane theory and residual learning in continuum mechanics to better situate the contribution.
  3. [Figures] Ensure all figures depicting learned anisotropy include quantitative measures (e.g., anisotropy ratios or eigenvalue spreads) rather than relying solely on visual inspection.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the opportunity to respond to the referee's report. We are grateful for the positive evaluation of the significance of our proposed MoSA framework and the detailed feedback provided. We address each of the major comments below and have revised the manuscript accordingly to incorporate the suggested improvements.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claims of 'superior accuracy, generalization, and robustness' and 'physically meaningful residual anisotropy' are presented without any quantitative error metrics, ablation results, baseline comparisons, or validation statistics. This absence is load-bearing for the central empirical claim that residual stress operators close the remaining real-to-sim gap after isotropic calibration.

    Authors: We agree that the abstract should include quantitative support for the claims to make them more compelling. Accordingly, we have revised the abstract to include specific quantitative metrics from our experiments, such as the percentage improvements in accuracy and generalization, references to ablation studies, and baseline comparisons. This revision ensures that the central empirical claims are substantiated directly in the abstract. revision: yes

  2. Referee: [Method] Method section on residual formulation: the weakest assumption—that mild residual anisotropy and heterogeneity dominate the post-calibration gap and can be captured without overfitting or violating consistency—requires explicit support. No analysis is visible showing that other sources (e.g., contact modeling, sensor noise, or discretization) are secondary, nor are there controls demonstrating that the learned operators remain physically admissible under the imposed motion constraints.

    Authors: We thank the referee for highlighting this important aspect. While the manuscript discusses the motivation for focusing on residual effects after calibration, we recognize the need for more explicit support. In the revised manuscript, we have added an analysis in the Method section addressing why other potential sources of the gap are secondary, based on our calibration diagnostics. We have also included additional experiments and controls to demonstrate that the learned residual operators maintain physical admissibility, including checks for consistency with the motion constraints and absence of overfitting through cross-validation. revision: yes

  3. Referee: [Experiments] Robot manipulation validation: the statement that better dynamics modeling 'translates into more reliable sim-to-real transfer' needs quantitative transfer metrics (success rates, trajectory error, or failure modes) with and without MoSA to establish the practical impact; qualitative demonstration alone is insufficient for the claim.

    Authors: We acknowledge that quantitative metrics would provide a more rigorous validation of the sim-to-real benefits. In the revised version, we have supplemented the robot manipulation experiments with quantitative results, including success rates for the manipulation tasks with and without MoSA, as well as trajectory error metrics and a discussion of observed failure modes. These additions directly address the need for measurable evidence of improved transfer reliability. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper's derivation starts from a standard isotropic continuum model as an external physics prior, then augments it with learned residual stress operators whose supervision comes from independent motion derivatives of the deformation field. No equation or claim reduces a reported prediction to a fitted parameter by construction, nor does any load-bearing step rely on a self-citation chain whose validity is internal to the present work. The residual formulation and microplane constraints are presented as additive corrections whose consistency is enforced by explicit derivative supervision rather than by redefinition of the target quantities. The framework is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that residuals after isotropic calibration are mild and learnable, plus fitted parameters in the residual network; no new entities are postulated.

free parameters (1)
  • residual stress operator weights
    Neural network parameters fitted to capture anisotropy and heterogeneity from visual observations.
axioms (1)
  • domain assumption Real-world objects exhibit only mild anisotropy and heterogeneity after the isotropic backbone is calibrated.
    Explicitly stated as the key bottleneck for further closing the real-to-sim gap.

pith-pipeline@v0.9.0 · 5806 in / 1214 out tokens · 38307 ms · 2026-05-22T06:54:01.706792+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    MoSA uses an isotropic model as a physics prior and learns residual stress operators to capture mild anisotropy and heterogeneity. It progressively adapts stresses via microplane-constrained redistribution in a physics-informed cascaded network.

  • IndisputableMonolith/Foundation/BranchSelection branch_selection unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    We keep a calibrated near-isotropic backbone to explain the dominant behavior, and learn only the residual effects that arise from mild anisotropy and heterogeneity.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    Differentiable material point method for the control of deformable objects.arXiv preprint arXiv:2512.13214,

    Bolliger, D., Fadini, G., Bambach, M., and Rupenyan, A. Differentiable material point method for the control of deformable objects.arXiv preprint arXiv:2512.13214,

    work page arXiv

  2. [2]

    Gic: Gaussian-informed continuum for physical property identification and simulation.arXiv preprint arXiv:2406.14927,

    Cai, J., Yang, Y ., Yuan, W., He, Y ., Dong, Z., Bo, L., Cheng, H., and Chen, Q. Gic: Gaussian-informed continuum for physical property identification and simulation.arXiv preprint arXiv:2406.14927,

    work page arXiv

  3. [3]

    Vid2sim: Generalizable, video-based reconstruction of appearance, geometry and physics for mesh-free simulation

    Chen, C., Dou, Z., Wang, C., Huang, Y ., Chen, A., Feng, Q., Gu, J., and Liu, L. Vid2sim: Generalizable, video-based reconstruction of appearance, geometry and physics for mesh-free simulation. InProceedings of the Computer Vision and Pattern Recognition Conference, pp. 26545– 26555, 2025a. Chen, G., Suri, S., Wu, Y ., V oulga, E., Levin, D. I., and Pai, ...

    work page arXiv

  4. [4]

    F., Chen, A

    Dagli, R., Xiang, D., Modi, V ., Loop, C., Tsang, C. F., Chen, A. H., Hu, A., State, G., Levin, D. I., and Shugrina, M. V omp: Predicting volumetric mechanical property fields. arXiv preprint arXiv:2510.22975,

    work page arXiv

  5. [5]

    Gaussianflow: Splatting gaussian dynamics for 4d content creation.arXiv preprint arXiv:2403.12365,

    Gao, Q., Xu, Q., Cao, Z., Mildenhall, B., Ma, W., Chen, L., Tang, D., and Neumann, U. Gaussianflow: Splatting gaussian dynamics for 4d content creation.arXiv preprint arXiv:2403.12365,

    work page arXiv

  6. [6]

    Physics-as- inverse-graphics: Unsupervised physical parameter es- timation from video.arXiv preprint arXiv:1905.11169,

    Jaques, M., Burke, M., and Hospedales, T. Physics-as- inverse-graphics: Unsupervised physical parameter es- timation from video.arXiv preprint arXiv:1905.11169,

    work page arXiv 1905

  7. [7]

    The material point method for simulating continuum materials

    Jiang, C., Schroeder, C., Teran, J., Stomakhin, A., and Selle, A. The material point method for simulating continuum materials. InAcm siggraph 2016 courses, pp. 1–52. ACM,

    work page 2016

  8. [8]

    Phys- twin: Physics-informed reconstruction and simulation of deformable objects from videos.arXiv preprint arXiv:2503.17973, 2025

    Jiang, H., Hsu, H.-Y ., Zhang, K., Yu, H.-N., Wang, S., and Li, Y . Phystwin: Physics-informed reconstruction and simulation of deformable objects from videos.arXiv preprint arXiv:2503.17973,

    work page arXiv

  9. [9]

    K., Ye, J., Duan, Y ., Abbeel, P., Wang, X., and Yi, S

    Jing, C., Bandi, J. K., Ye, J., Duan, Y ., Abbeel, P., Wang, X., and Yi, S. Contact-aware neural dynamics.arXiv preprint arXiv:2601.12796,

    work page arXiv

  10. [10]

    Li, Y .-L

    URL https://repo-sam.inria.fr/ fungraph/3d-gaussian-splatting/. Li, X., Qiao, Y .-L., Chen, P. Y ., Jatavallabhula, K. M., Lin, M., Jiang, C., and Gan, C. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification.arXiv preprint arXiv:2303.05512,

    work page arXiv

  11. [11]

    Omniphysgs: 3d con- stitutive gaussians for general physics-based dynamics generation.arXiv preprint arXiv:2501.18982,

    Lin, Y ., Lin, C., Xu, J., and Mu, Y . Omniphysgs: 3d con- stitutive gaussians for general physics-based dynamics generation.arXiv preprint arXiv:2501.18982,

    work page arXiv

  12. [12]

    Gaussian- augmented physics simulation and system identification with complex colliders.arXiv preprint arXiv:2511.06846,

    Vasile, F., Qiu, R.-Z., Natale, L., and Wang, X. Gaussian- augmented physics simulation and system identification with complex colliders.arXiv preprint arXiv:2511.06846,

    work page arXiv

  13. [13]

    Neuspring: Neural spring fields for reconstruction and simulation of deformable objects from videos.arXiv preprint arXiv:2511.08310, 2025a

    Xu, Q., Liu, J., Yu, S., Wang, Y ., Zhou, Y ., Zhou, J., Cui, J., Ong, Y .-S., and Zhang, H. Neuspring: Neural spring fields for reconstruction and simulation of deformable objects from videos.arXiv preprint arXiv:2511.08310, 2025a. Xu, X., Ge, W., Qiu, D., Chen, Z., Yan, D., Liu, Z., Zhao, H., Zhao, H., Zhang, S., Liang, J., and Chen, Y .-C. Gaussian- pr...

    work page arXiv

  14. [14]

    Adaptigraph: Material-adaptive graph-based neural dynamics for robotic manipulation.arXiv preprint arXiv:2407.07889,

    Zhang, K., Li, B., Hauser, K., and Li, Y . Adaptigraph: Material-adaptive graph-based neural dynamics for robotic manipulation.arXiv preprint arXiv:2407.07889,

    work page arXiv

  15. [15]

    R., and Martius, G

    Zhobro, M., Geist, A. R., and Martius, G. Learning 3d- gaussian simulators from rgb videos.arXiv preprint arXiv:2503.24009,

    work page arXiv