Pith. sign in

REVIEW 2 major objections 5 minor 31 references

A differential-amplifier-inspired attention mechanism suppresses noise in multi-view robot vision, raising success rates while cutting training time by a third.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 06:38 UTC pith:WT5FLCFL

load-bearing objection Solid multi-view robotics result with a real efficiency edge; the differential-amplifier story is mostly rebranded Diff Transformer plus a small hinge loss, and the ablations never fully isolate it. the 2 major comments →

arxiv 2607.02845 v1 pith:WT5FLCFL submitted 2026-07-03 cs.RO cs.AIcs.CV

Differential Amplifier-Inspired AmpAttention for Multi-View Robotic Manipulation

Jin Yang , Ping Wei , Nanning Zheng This is my paper
classification cs.RO cs.AIcs.CV
keywords AmpAttentionmulti-view robotic manipulationattention driftdifferential amplifierCMRR lossRLBenchRVAFSAM2
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Multi-view robotic manipulation often fails when standard attention drifts toward background clutter, occlusions, or repeated patterns instead of the objects that matter. The authors treat this as an analog-circuit problem: they design AmpAttention so that it amplifies the difference between two attention branches while a CMRR-style loss penalizes the shared (common-mode) noise. Placed first as task-guided intra-view attention and then as inter-view fusion, the mechanism forms the RVAF model. On 18 RLBench tasks covering 249 variations, RVAF records the highest average success among the methods compared and finishes training in roughly two-thirds the time of a re-run of the prior state-of-the-art on identical hardware. Adding a large pretrained image encoder (RVAF++) further lifts high-precision assembly tasks, reaching 91 percent success on insert-peg, and the same architecture works from a single third-person camera on real-robot dart insertion and switch-pressing after only fifty demonstrations.

Core claim

AmpAttention, by computing a learnable differential of two softmax attentions and regularizing it with a hinge common-mode-rejection loss, produces higher signal-to-noise visual cues than standard multi-head attention; embedding this operator as task-guided intra-view then inter-view blocks yields a multi-view policy (RVAF) that simultaneously improves average success and reduces training cost on a large suite of manipulation tasks.

What carries the argument

AmpAttention: split query/key/value into two groups, form differential attention A_diff = A2 − λ A1 and common-mode A_cm = (A1 + A2)/2, then train with a hinge CMRR loss that penalizes cases where the common-mode norm is not smaller than the differential norm; the resulting heads replace ordinary self- and cross-attention throughout the network.

Load-bearing premise

The paper assumes that attention drift caused by common-mode visual noise is the main performance bottleneck, and that maximizing differential over common-mode attention is what actually drives the measured gains.

What would settle it

An ablation that keeps every architectural detail of RVAF identical but replaces AmpAttention with ordinary multi-head attention (or removes only the CMRR loss) and shows no drop in average success rate or TRR–TIR ratio on the same 18 RLBench tasks would falsify the central causal claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes AmpAttention, an attention operator inspired by differential amplifiers, that splits Q/K/V into two branches, forms a differential attention A_diff = A2 − λ A1 and a common-mode term A_cm, and regularizes them with a hinge CMRR-style loss L_CMRR (Eqs. 5–7, §III.C). This is embedded in RVAF, which applies task-guided intra-view AmpAttention followed by inter-view AmpAttention on multi-view orthographic images, and in RVAF++ which freezes a LoRA-adapted SAM2 image encoder. On 18 RLBench tasks (249 variations) RVAF reports 83.1 % average success in 0.24 days on 4090s versus a same-hardware RVT-2* baseline of 80.3 % in 0.36 days; RVAF++ reaches 87.0 % overall and 91 % on insert peg. Ablations (Table II), four-task VLA comparisons, five real-robot tasks (10 trials each), and attention heatmaps are provided.

Significance. If the gains hold, the work supplies a practical, training-efficient multi-view architecture that improves both average success and high-precision tasks while cutting wall-clock time by roughly one-third relative to a re-trained RVT-2. The same-hardware re-baseline, multi-seed planner evaluation, component ablations, and real-robot transfer (including dart-to-bullseye insertion) are concrete engineering contributions that the community can build on. The differential-amplifier framing and CMRR loss are a novel inductive bias for robotic attention, even if their causal necessity is not fully isolated.

major comments (2)
  1. §III.C and Table II (rows 1–3): AmpAttention is explicitly built on the Differential Transformer dual-softmax construction (citation [24], λ parameterization, head scaling). The only ablation contrasts full AmpAttn (83.1 %) vs AmpAttn without L_CMRR (81.6 %) vs standard attention (81.2 %). There is no head-to-head against the plain dual-branch residual of [24]. Consequently the modest +1.5–1.9 % cannot be attributed specifically to the circuit analogy or to common-mode rejection rather than to any dual-branch design. A controlled comparison (or an explicit statement that the contribution is the robotic packaging of [24] plus L_CMRR) is needed for the mechanism claim that underpins every AmpAttention block.
  2. §I and Fig. 1: The central narrative asserts that “attention drift” is the dominant failure mode of multi-view robotic attention and that maximizing the TRR–TIR ratio via AmpAttention is what drives the success-rate gains. Fig. 1’s ratios are post-hoc visualizations on a single scene; no quantitative correlation between TRR–TIR improvement and task success across the 18 tasks, nor an intervention that deliberately injects or removes attention noise, is reported. Without such evidence the causal story remains under-supported relative to the architectural and training-schedule differences that also exist between RVAF and RVT-2*.
minor comments (5)
  1. Table I: report absolute wall-clock hours (or GPU-hours) alongside the “days” figures so that the 33.3 % claim is reproducible across hardware.
  2. Eq. (6): the hinge max(0, log(1+∥Acm∥) − log(1+∥Adiff∥)) is never ablated against a soft ratio or an L2 penalty; a one-sentence sensitivity note would strengthen the design choice.
  3. §IV.C: the drop on “slide blocks” for RVAF++ is attributed to multi-task trade-off; a brief per-task learning-curve or capacity analysis would make the claim less speculative.
  4. Fig. 7 heatmaps: add a quantitative TRR/TIR or entropy metric next to the qualitative images so that the visual claim can be checked.
  5. Notation: λ and λ_init appear in both the differential operator and the head-scaling formula; a short glossary or consistent subscripting would reduce reader friction.

Circularity Check

0 steps flagged

No circularity: AmpAttention is an architectural proposal with independent held-out success-rate evaluation; L_CMRR regularizes internals but does not algebraically define task metrics.

full rationale

This is a standard empirical robotics/ML paper. AmpAttention (Eqs. 3–7) and the hinge L_CMRR (Eq. 6) are design choices inspired by differential amplifiers and the Differential Transformer parameterization (external citation [24]); they are optimized jointly with action losses (Eq. 9) during training. Reported claims (Table I average success 83.1 % / 87.0 %, 33.3 % wall-clock reduction vs re-trained RVT-2*, 91 % on insert peg, real-world dart insertion) are measured on held-out RLBench episodes (25 demos/task) and real-robot trials, not recovered by construction from the CMRR objective or from any fitted parameter. Ablations (Table II) and attention visualizations (Fig. 1) are post-hoc diagnostics, not load-bearing uniqueness or self-definitional steps. Inheritance of multi-view rendering and key-frame prediction from RVT/RVT-2 is ordinary architectural reuse, not a circular derivation of the performance numbers. No self-definitional loop, no fitted-input-as-prediction, no load-bearing self-citation uniqueness theorem, and no renaming of a known result as a first-principles derivation appear in the chain.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 3 invented entities

The paper is an empirical systems contribution. Load-bearing free parameters are loss weights, architectural depths, and rendering choices tuned for RLBench. Domain axioms inherit the RVT multi-view key-frame pipeline. Invented pieces are AmpAttention (as packaged), the CMRR attention loss, and the RVAF models; independent evidence is only the paper’s own ablations and success rates, not external theory.

free parameters (6)
  • α2 (CMRR loss weight) = 0.01
    Set to 0.01 by hand in Eq. 9; balances action CE losses against CMRR. Central claim that CMRR improves focus depends on this scale.
  • λ / λ_init (differential branch balance) = learnable; λ_init in (0,1)
    Learnable λ with λ_init scaling of heads, following [24]; controls how strongly common-mode is subtracted. Affects AmpAttention behavior.
  • L (AmpAttention module depth) = 4
    Stack depth for intra- and inter-view blocks set to 4; capacity hyperparameter not derived.
  • LoRA rank for SAM2 encoder = 16
    Default rank 16 for RVAF++; controls adaptation capacity of the foundation encoder.
  • fine-stage zoom-in scale and # of virtual views = scale=4, views=3
    Chosen via ablation (scale 4, 3 views); performance and training time depend on these rendering choices.
  • training schedule (steps, batch, LR) = as stated in §IV.A
    RVAF ~50K steps, bs 160, LR 2e-3; RVAF++ ~110K steps, bs 64, LR 3.2e-3. Results are schedule-dependent.
axioms (4)
  • domain assumption Multi-view orthographic images re-rendered from reconstructed point clouds (RVT-style) plus key-frame action labels are a sufficient observation/action interface for the claimed tasks.
    Inherited from RVT/RVT-2; entire architecture and evaluation rest on this pipeline (§III.B, §IV.A).
  • ad hoc to paper Differential-amplifier common-mode rejection is a valid model of “attention noise” in robotic multi-view images, so maximizing differential over common-mode attention outputs improves task success.
    Core design hypothesis in §I and §III.C; not proven outside the paper’s ablations.
  • domain assumption Standard cross-entropy on translation, rotation, gripper, and collision heads is an adequate action objective when combined with small CMRR regularization.
    Follows prior RVT-family training (§III.G).
  • standard math Softmax attention and multi-head residual Transformer blocks are the right substrate for injecting the differential operator.
    Standard Transformer machinery used throughout §III.
invented entities (3)
  • AmpAttention no independent evidence
    purpose: Attention operator that computes differential and common-mode branches and is trained with CMRR loss to suppress attention noise.
    Packaged as the paper’s main mechanism; closely related to Differential Transformer but with CMRR loss and robotics-specific use. No independent evidence outside this work’s metrics.
  • CMRR loss for attention (L_CMRR) no independent evidence
    purpose: Hinge on log norms of common-mode vs differential attention outputs to encourage high signal-to-noise attention maps.
    Defined in Eq. 6; unique to this paper’s training objective. Validated only via internal ablations.
  • RVAF / RVAF++ no independent evidence
    purpose: Full multi-view manipulation models combining task-guided intra-view and inter-view AmpAttention (plus SAM2 in ++).
    System-level contribution; performance claims are the paper’s empirical results.

pith-pipeline@v1.1.0-grok45 · 18381 in / 3705 out tokens · 42126 ms · 2026-07-12T06:38:36.360812+00:00 · methodology

0 comments
read the original abstract

Multi-view robotic manipulation methods with the attention mechanism have recently achieved significant progress in both training efficiency and task performance. However, the inherent redundancy, occlusion, and viewpoint dependency in robotic view images often lead to severe attention drift. To address this challenge, we propose AmpAttention, a novel attention mechanism inspired by differential amplifiers in analog circuits. It aims to suppress attention noise and capture high signal-to-noise ratio signals for more reliable perception. Based on this, we introduce the RVAF model, which integrates task-guided intra-view and inter-view AmpAttention. Compared to previous state-of-the-art methods, RVAF achieves the optimal average success rate across 18 RLBench tasks (249 variations) while reducing training time by 33.3\%. RVAF also demonstrates strong potential in real-world high-precision tasks, exemplified by its ability to pick up a dart and accurately insert it into the red bullseye. Furthermore, we extend RVAF to RVAF++ by incorporating the SAM2 image encoder. RVAF++ achieves substantial gains on high-precision tasks, achieving a 91\% success rate on the `insert peg' task. More qualitative results are provided at the anonymous project website https://anonymous.4open.science/w/RVAF-Anonymization.

Figures

Figures reproduced from arXiv: 2607.02845 by Jin Yang, Nanning Zheng, Ping Wei.

Figure 1
Figure 1. Figure 1: Comparison of attention distributions between standard attention [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the differential amplifier. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the AmpAttention. based on the predicted position. These features are con￾catenated and processed by the MLP to predict the gripper orientation, collision state, and gripper state. C. AmpAttention Existing view-based robotic manipulation methods rely on standard attention to extract task-relevant cues [2], [5]. However, the complexity and variability of manipulation sce￾narios present significa… view at source ↗
Figure 5
Figure 5. Figure 5: Real-world experimental setup. (a) Real-world task scenario setting. (b) Pick dart and insert bullseye: the red/green bullseye has a radius of 6/14 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Robustness validation of RVAF under the real-world environment. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of attention heatmaps for different models across various [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

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

31 extracted references · 18 linked inside Pith

  1. [1]

    Perceiver-actor: A multi- task transformer for robotic manipulation,

    M. Shridhar, L. Manuelli, and D. Fox, “Perceiver-actor: A multi- task transformer for robotic manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 785–799

    2023

  2. [2]

    Rvt: Robotic view transformer for 3d object manipulation,

    A. Goyal, J. Xu, Y . Guo, V . Blukis, Y .-W. Chao, and D. Fox, “Rvt: Robotic view transformer for 3d object manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 694–710

    2023

  3. [3]

    Rvt- 2: Learning precise manipulation from few demonstrations,

    A. Goyal, V . Blukis, J. Xu, Y . Guo, Y .-W. Chao, and D. Fox, “Rvt- 2: Learning precise manipulation from few demonstrations,”arXiv preprint arXiv:2406.08545, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Sam2act: Integrating visual foundation model with a memory architecture for robotic manipulation,

    H. Fang, M. Grotz, W. Pumacay, Y . R. Wang, D. Fox, R. Krishna, and J. Duan, “Sam2act: Integrating visual foundation model with a memory architecture for robotic manipulation,”arXiv preprint arXiv:2501.18564, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Sam-e: Leveraging visual foundation model with sequence imitation for embodied manipulation,

    J. Zhang, C. Bai, H. He, Z. Wang, B. Zhao, X. Li, and X. Li, “Sam-e: Leveraging visual foundation model with sequence imitation for embodied manipulation,” inInternational Conference on Machine Learning. PMLR, 2024, pp. 58 579–58 598

    2024

  6. [6]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafsonet al., “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Mastering visual continuous control: Improved data-augmented reinforcement learning,

    D. Yarats, R. Fergus, A. Lazaric, and L. Pinto, “Mastering visual continuous control: Improved data-augmented reinforcement learning,” arXiv preprint arXiv:2107.09645, 2021

    Pith/arXiv arXiv 2021

  8. [8]

    Q-attention: Enabling efficient learning for vision-based robotic manipulation,

    S. James and A. J. Davison, “Q-attention: Enabling efficient learning for vision-based robotic manipulation,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 1612–1619, 2022

    2022

  9. [9]

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation,

    Q. Li, Y . Liang, Z. Wang, L. Luo, X. Chen, M. Liao, F. Wei, Y . Deng, S. Xu, Y . Zhanget al., “Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation,” arXiv preprint arXiv:2411.19650, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Open- vla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketiet al., “Open- vla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Tinyvla: Towards fast, data-efficient vision- language-action models for robotic manipulation,

    J. Wen, Y . Zhu, J. Li, M. Zhu, Z. Tang, K. Wu, Z. Xu, N. Liu, R. Cheng, C. Shenet al., “Tinyvla: Towards fast, data-efficient vision- language-action models for robotic manipulation,”IEEE Robotics and Automation Letters, 2025

    2025

  12. [12]

    Fourier transporter: Bi-equivariant robotic manipulation in 3d,

    H. Huang, O. Howell, D. Wang, X. Zhu, R. Walters, and R. Platt, “Fourier transporter: Bi-equivariant robotic manipulation in 3d,”arXiv preprint arXiv:2401.12046, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Act3d: 3d feature field transformers for multi-task robotic manipulation,

    T. Gervet, Z. Xian, N. Gkanatsios, and K. Fragkiadaki, “Act3d: 3d feature field transformers for multi-task robotic manipulation,”arXiv preprint arXiv:2306.17817, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Assistive tele-op: Leveraging transformers to collect robotic task demonstrations,

    H. M. Clever, A. Handa, H. Mazhar, K. Parker, O. Shapira, Q. Wan, Y . Narang, I. Akinola, M. Cakmak, and D. Fox, “Assistive tele-op: Leveraging transformers to collect robotic task demonstrations,”arXiv preprint arXiv:2112.05129, 2021

    Pith/arXiv arXiv 2021

  15. [15]

    The ingredients for robotic diffusion transformers,

    S. Dasari, O. Mees, S. Zhao, M. K. Srirama, and S. Levine, “The ingredients for robotic diffusion transformers,”arXiv preprint arXiv:2410.10088, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Octo: An open-source generalist robot policy,

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xuet al., “Octo: An open-source generalist robot policy,”arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Vision-language-action model with speech instructions for customized robot manipulation,

    W. Zhao, P. Ding, M. Zhang, Z. Gong, S. Bai, H. Zhao, and D. Wang, “Vision-language-action model with speech instructions for customized robot manipulation,”arXiv preprint arXiv:2502.13508, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Manipllm: Embodied multimodal large language model for object-centric robotic manipulation,

    X. Li, M. Zhang, Y . Geng, H. Geng, Y . Long, Y . Shen, R. Zhang, J. Liu, and H. Dong, “Manipllm: Embodied multimodal large language model for object-centric robotic manipulation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18 061–18 070

    2024

  19. [19]

    Progprompt: Generating situated robot task plans using large language models,

    I. Singh, V . Blukis, A. Mousavian, A. Goyal, D. Xu, J. Tremblay, D. Fox, J. Thomason, and A. Garg, “Progprompt: Generating situated robot task plans using large language models,” in2023 IEEE Interna- tional Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 11 523–11 530

    2023

  20. [20]

    Instruction-driven history-aware policies for robotic ma- nipulations,

    P.-L. Guhur, S. Chen, R. G. Pinel, M. Tapaswi, I. Laptev, and C. Schmid, “Instruction-driven history-aware policies for robotic ma- nipulations,” inConference on Robot Learning. PMLR, 2023, pp. 175–187

    2023

  21. [21]

    W. M. Sansen,Analog design essentials. Springer Science & Business Media, 2007, vol. 859

    2007

  22. [22]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PMLR, 2021, pp. 8748–8763

    2021

  23. [23]

    Flashattention-2: Faster attention with better parallelism and work partitioning,

    T. Dao, “Flashattention-2: Faster attention with better parallelism and work partitioning,”arXiv preprint arXiv:2307.08691, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Differential transformer,

    T. Ye, L. Dong, Y . Xia, Y . Sun, Y . Zhu, G. Huang, and F. Wei, “Differential transformer,”arXiv preprint arXiv:2410.05258, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Glu variants improve transformer,

    N. Shazeer, “Glu variants improve transformer,”arXiv preprint arXiv:2002.05202, 2020

    Pith/arXiv arXiv 2002

  26. [26]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  27. [27]

    Rlbench: The robot learning benchmark & learning environment,

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison, “Rlbench: The robot learning benchmark & learning environment,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3019–3026, 2020

    2020

  28. [28]

    Large batch optimization for deep learning: Training bert in 76 minutes,

    Y . You, J. Li, S. Reddi, J. Hseu, S. Kumar, S. Bhojanapalli, X. Song, J. Demmel, K. Keutzer, and C.-J. Hsieh, “Large batch optimization for deep learning: Training bert in 76 minutes,”arXiv preprint arXiv:1904.00962, 2019

    Pith/arXiv arXiv 1904

  29. [29]

    Coarse-to- fine q-attention: Efficient learning for visual robotic manipulation via discretisation,

    S. James, K. Wada, T. Laidlow, and A. J. Davison, “Coarse-to- fine q-attention: Efficient learning for visual robotic manipulation via discretisation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 13 739–13 748

    2022

  30. [30]

    π 0: A vision- language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichteret al., “π 0: A vision- language-action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Fine-tuning vision-language- action models: Optimizing speed and success,

    M. J. Kim, C. Finn, and P. Liang, “Fine-tuning vision-language- action models: Optimizing speed and success,”arXiv preprint arXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025