Pith. sign in

REVIEW 3 major objections 7 minor 43 references

Frozen robot models learn new tasks from raw human video at test time

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 · glm-5.2

2026-07-09 22:05 UTC pith:6VH44RPU

load-bearing objection Solid new idea with one unfair baseline comparison that doesn't sink the core contribution the 3 major comments →

arxiv 2607.06988 v1 pith:6VH44RPU submitted 2026-07-08 cs.RO cs.AI

WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time

Yusen Feng , Bingchen Han , Jiangran Lyu , Kai Liu , Yixin Zheng , Yuxuan Wan , Weiheng Liu , Sun Han
show 8 more authors
Ruiqin Li Yulong Zhang Fangfu Liu Xuesong Shi Libin Liu Yizhou Wang Zhizheng Zhang He Wang
This is my paper
classification cs.RO cs.AI
keywords test-time trainingworld-action modelhuman video learningrobot foundation modelfast-weight memorykey-value memory reconstructionmanipulationtest-time adaptation
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.

The paper proposes a method called WAM-TTT for adapting a pretrained world-action model (a robot foundation model that jointly predicts visual futures and generates actions) to new task variants at deployment time, using nothing more than unlabeled egocentric human video. The central mechanism is a lightweight fast-weight memory module attached to the model's video-processing pathway. During a meta-training stage, this memory is calibrated on paired human-robot demonstrations so that human visual cues become aligned with executable robot behaviors through a key-value memory reconstruction objective. At test time, the pretrained model stays frozen; only the memory module updates, via self-supervised video prediction on the human clips. The adapted memory then steers action generation through the model's shared video-action representation. The paper claims this approach consistently outperforms feeding the same human videos as in-context conditioning tokens, across nine manipulation tasks on three robot embodiments, in both standardized and unseen household environments.

Core claim

The paper claims that absorbing human videos into a parametric fast-weight memory through self-supervised video prediction is a more effective and robust way to transfer skill from human demonstration to a robot policy than feeding those same videos as in-context conditioning tokens. The key-value memory reconstruction loss is shown analytically (in the linear case) to produce a weight matrix equivalent to linear attention against the human key-value cache, meaning the inner-loop optimization directly constructs an attention-like retrieval mechanism from human visual cues to robot queries without requiring any explicit action labels, hand poses, or retargeted trajectories on the human side.

What carries the argument

The method has three load-bearing components. First, a world-action model (WAM) that jointly models visual dynamics and robot actions through coupled video and action experts in a diffusion transformer. Second, a test-time training (TTT) layer: a fast-weight memory branch added to the video expert that updates via inner-loop SGD during each forward pass. Third, a key-value memory reconstruction loss that trains the fast-weight network to map human-derived keys to human-derived values, which analytically reduces to a linear-attention readout when the robot side queries it. The meta-training stage optimizes the slow projections and initializations of this interface on paired human-robot data,;

Load-bearing premise

The meta-training stage assumes that phase-aligned paired human-robot data is sufficient to learn a human-to-robot memory interface that generalizes to unseen tasks and environments at deployment, and the paper acknowledges that the boundary of this generalization has not been characterized empirically.

What would settle it

If the fast-weight memory, after meta-training, fails to produce useful control signals when fed human videos from tasks or visual domains sufficiently different from the meta-training distribution, or if the key-value memory reconstruction loss does not produce attention-like retrieval behavior in the nonlinear MLP case, the central mechanism would not hold.

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

If this is right

  • A robot deployed in a new home could be shown a task once via a head-mounted camera by a non-expert, and the policy would adapt without any fine-tuning of the core model or any robot-side demonstration collection.
  • The parametric memory approach sidesteps the growing context-length problem of in-context learning, since each new demonstration updates a fixed-size set of fast weights rather than appending tokens to a context window.
  • The decoupling of human-side adaptation from the frozen action prior suggests a path where foundation models ship with explicit 'adaptation seats' that downstream users can drive with whatever side information they have.
  • The analytical equivalence between the memory reconstruction loss and linear attention provides a principled justification for why self-supervised video prediction on human data produces useful control signals: it is literally constructing a cross-attention retrieval mechanism.

Where Pith is reading between the lines

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

  • If the fast-weight memory can be driven by video prediction alone, it could in principle be driven by any modality with a phase-pairable training signal, such as audio of contact events or depth maps, opening a general interface for multi-modal test-time adaptation.
  • The 39-point gap between WAM-TTT and in-context conditioning on unseen tasks suggests that long-context conditioning on raw video may be fundamentally fragile under visual distribution shift, since perturbed visual statistics flow directly into the policy stream rather than being filtered through a learned memory interface.
  • The saturation of human-data benefit at 100 episodes per task, combined with the 1-for-1 substitution of human for robot data at equal total budget, implies that the bottleneck for robot learning may shift from data collection to the quality of the human-to-robot alignment interface itself.

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

3 major / 7 minor

Summary. The paper presents WAM-TTT, a test-time training framework for steering frozen world-action models (WAMs) using unlabeled egocentric human videos. The method attaches lightweight fast-weight (TTT) memory branches to the video expert of a pretrained WAM (LDA backbone). A meta-training stage uses paired human-robot data and a key-value memory reconstruction loss to calibrate the TTT interface so that human-video keys/values align with robot-side queries. At deployment, only the fast weights are updated via self-supervised video prediction on unseen human videos, while the WAM backbone and slow projections remain frozen. Experiments span three robot embodiments and nine manipulation tasks, with ablations isolating the TTT mechanism, meta-training, and memory reconstruction loss. The central claim is that absorbing human videos as fast-weight memory outperforms in-context video conditioning, co-training, and no-adaptation baselines, particularly under out-of-distribution household-scene perturbations.

Significance. The paper addresses a practically important problem: steering robot foundation models toward new task variants without collecting additional robot demonstrations or fine-tuning the full model. The idea of using action-free human videos as deployment-time memory for a frozen WAM is novel and well-motivated. The linear-attention witness analysis (Eq. A.2-A.4) provides a self-contained theoretical justification for why the key-value memory reconstruction loss induces attention-like retrieval behavior, which is a strength. The experimental scope—three embodiments, nine tasks, both in-distribution and OOD settings, plus ablations on data ratio, architecture, and pseudo-action injection—is commendable. The finding that retargeted pseudo-actions are net-negative (Table E.3, -43.4 pts) is a useful negative result for the community. The generalization preservation analysis (Table 3) and the six-axis perturbation study (Appendix E.3) provide additional evidence that the TTT mechanism does not catastrophically overwrite pretrained capabilities.

major comments (3)
  1. The WAM-ICL baseline comparison is structurally unfair and undermines the paper's headline claim. Section 4.2 states that 'the gap against WAM-ICL is the strongest piece of evidence for the design hypothesis,' and the abstract claims WAM-TTT 'consistently outperforms in-context human-video conditioning baselines.' However, WAM-ICL is defined only as 'the same WAM that ingests deployment-time human videos as in-context demonstrations, with no fast-weight adaptation' (Section 4.2). No implementation details are given for how human videos are tokenized and fed into the WAM's context, and critically, the paper itself acknowledges in Section 1 that in-context conditioning 'requires learning such capabilities during large-scale pretraining.' The WAM-ICL baseline receives no such pretraining, no paired human-robot meta-training, no TTT architecture, and no KVM objective. The symptoms confirm an
  2. unfair comparison: WAM-ICL scores 48.4% in Orig (below LDA's 50.2% with zero human data) and collapses to 7.1% in New (far below LDA's 32.5%). In-context human video actively hurts relative to no human data at all, which is not what a properly trained ICL baseline should do. The paper should either (a) provide a fair ICL baseline where the WAM is trained to consume in-context video tokens during pretraining or meta-training, or (b) substantially soften the claim of 'consistently outperforms in-context conditioning' to acknowledge that the comparison is against an untrained ICL interface. The independent evidence from the LDA comparison (+13.7 pts) and the ablations (Table 2) does support WAM-TTT's value, but the specific ICL comparison as currently constructed is not load-bearing evidence.
  3. The generalization boundary of the meta-training interface is uncharacterized and limits the interpretability of the results. Section 5 states that 'the further the deployment task drifts from the meta-training pairing distribution, the weaker the adaptation' and that this boundary 'has not been characterised empirically.' All nine evaluation tasks are also represented in the meta-training dataset (Section 4.1: 2,286 paired episodes covering 9 tasks). The paper does not test on a task that is entirely absent from meta-training, so it is unclear whether the TTT mechanism generalizes to genuinely novel skills or merely adapts to new visual conditions of known skills. The 'New' setting changes lighting, objects, and table height but preserves task structure. Adding at least one held-out task (absent from meta-training) to the evaluation would substantially strengthen the claim of 'steering'
minor comments (7)
  1. Appendix D, Stamp Paper rubric: 'stamp s[uccessfully grasped' contains a stray bracket and broken word.
  2. Table 3 caption: 'Diliver Drink' should be 'Deliver Drink'.
  3. Section 4.1: the progress metric is defined only briefly as 'partial-credit' with reference to 'recent VLA evaluations.' The full rubric in Appendix D should be referenced here, and the potential subjectivity of task-specific additive scoring should be acknowledged.
  4. Table B.1: N=1 inner SGD iteration for both meta-training and test-time is a surprisingly small value. A brief justification or sensitivity analysis would strengthen the paper.
  5. Section 3.2, Eq. (3): the normalization constant BL_h d is described as making the loss 'invariant to mini-batch size, human-sequence length, and embedding dimension.' This is correct for the linear case but the justification for the nonlinear MLP case is not provided. A brief note would help.
  6. Figure 2: the diagram is dense and the distinction between meta-training and test-time pathways could be more clearly separated visually. The current single-figure layout conflates two stages that are described separately in the text.
  7. The paper uses 'WAM' and 'WAM-TTT' interchangeably in some places (e.g., Section 3.1 title 'World Action Model' vs. the method name). Consistent capitalization would improve readability.

Circularity Check

0 steps flagged

No circularity found: the derivation chain is self-contained, with no prediction reducing to its inputs by construction.

full rationale

I walked the paper's full derivation chain and found no circularity. (1) The core mechanism—meta-train TTT fast weights on paired human-robot data (Eq. 4–6), then adapt only fast weights at test time on human video (Eq. 7–8)—is a design with independent inputs (paired data, video prediction loss, KVM loss) and independent evaluation (robot rollout progress). No step defines a quantity in terms of the result it claims to derive. (2) The linear-attention witness (Eq. A.2–A.4) is a self-contained mathematical argument: minimizing the KVM loss in the linear special case yields W* = V_h^T K_h (K_h^T K_h)^{-1}, which under a stated isotropy hypothesis reduces to the Hebbian outer product, and querying with robot Q_r produces linear-attention readout. This derivation does not assume its conclusion. (3) The LDA backbone [22] has overlapping authors but is a tool/architecture dependency (like citing a framework), not a logical step where the result is defined in terms of itself. Spatial-TTT [34], the other load-bearing methodological citation, has no author overlap with this paper. (4) The WAM-ICL baseline comparison flagged by the skeptic is a fairness/correctness concern (the baseline may not have been trained to handle in-context video), not circularity: the paper does not define WAM-ICL's results in terms of WAM-TTT's, nor does any equation reduce one to the other. The paper also compares against fully external baselines (π0.5, EGOSCALE) and runs ablations (Table 2, Table E.1–E.3) that provide independent support. No step in the derivation chain reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

The free parameters are standard hyperparameters tuned on the training set. The axioms are domain assumptions typical for robot learning from demonstration. The invented entities are the core methodological contributions, which are empirically validated rather than postulated without evidence.

free parameters (4)
  • lambda (memory reconstruction weight) = 4e-2
    Weighting factor for the KVM loss, chosen empirically.
  • N (inner SGD iterations) = 1
    Number of inner-loop gradient steps, set to 1 for both meta-training and test-time.
  • eta (inner LR) = 0.1 (meta), 0.01 (test)
    Learning rate for the fast-weight inner SGD updates.
  • TTT head dim d / fast-weight hidden width f_h = 48 / 128
    Architectural dimensions for the TTT layer.
axioms (3)
  • domain assumption Phase alignment between human and robot data is a sufficient proxy for temporal synchronization.
    Section 3.2 assumes normalized phase matching is adequate for meta-training alignment.
  • domain assumption The linear-attention witness (Eq. A.2-A.4) generalizes to the nonlinear MLP case.
    Appendix A uses the linear case to justify the KVM loss behavior, assuming the MLP acts as a smooth analog.
  • domain assumption The LDA backbone provides a shared visual-action representation that can be steered by video-side residuals.
    The entire method relies on the WAM having a coupled video-action representation where video residuals affect action generation.
invented entities (2)
  • TTT fast-weight memory branch independent evidence
    purpose: To store human video information in parametric weights that can be queried by robot tokens.
    The entity is the core contribution; its utility is tested via ablation (Table 2) and it makes falsifiable predictions about performance.
  • Key-value memory reconstruction loss (L_KVM) independent evidence
    purpose: To train the fast-weight network to act like a cross-attention mechanism from robot queries to human keys/values.
    The loss is defined and its effect is measured in ablations (w/o Memory Recon.).

pith-pipeline@v1.1.0-glm · 25472 in / 2286 out tokens · 279413 ms · 2026-07-09T22:05:31.076500+00:00 · methodology

0 comments
read the original abstract

Steering robot foundation models (RFMs) toward new task variants or user-preferred behaviors remains challenging, often requiring additional robot demonstrations, task-specific fine-tuning, or long-context conditioning. We present WAM-TTT, a test-time training framework for steering world action models from raw human videos. Rather than treating human videos as trajectories to imitate, WAM-TTT absorbs them into a lightweight adaptive memory inside a frozen WAM through self-supervised video prediction. To make this memory useful for control, we introduce a meta-training stage that aligns human demonstrations with robot behaviors using paired human-robot data and a key--value memory reconstruction objective. At test time, only unlabeled human videos are required to adapt the memory, while the pretrained WAM remains frozen. This enables efficient and reusable steering without robot actions, human-side annotations, or task-specific fine-tuning, while preserving the generalization ability of the foundation model. Extensive experiments show that WAM-TTT consistently outperforms in-context human-video conditioning baselines across diverse manipulation tasks and generalization settings.

Figures

Figures reproduced from arXiv: 2607.06988 by Bingchen Han, Fangfu Liu, He Wang, Jiangran Lyu, Kai Liu, Libin Liu, Ruiqin Li, Sun Han, Weiheng Liu, Xuesong Shi, Yixin Zheng, Yizhou Wang, Yulong Zhang, Yusen Feng, Yuxuan Wan, Zhizheng Zhang.

Figure 1
Figure 1. Figure 1: Overview of WAM-TTT. Given unlabeled human demonstrations from diverse environ￾ments, WAM-TTT steers a pretrained World Action Model (WAM) without retargeting, robot actions, or human-side annotations. During deployment, human videos are absorbed into lightweight TTT fast weights through self-supervised video prediction, while the pretrained action model remains frozen. The adapted memory then guides robot… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of WAM-TTT. We first meta-train a fast-weight memory using paired human-robot demonstrations, encouraging human visual cues to align with robot behaviors through a key–value memory reconstruction objective. At test time, the memory is adapted from unlabeled human videos via video prediction, while the pretrained WAM remains frozen. The adapted memory then steers robot execution through the WAM’s s… view at source ↗
Figure 3
Figure 3. Figure 3: Experimental setup [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative rollouts. For each unseen task we show a robot rollout filmstrip (right) and the paired human demonstration used as deployment-time Key/Value (left). Dataset and Metric. We collect a meta-training dataset consisting of 2,286 paired human and robot episodes, which broadly covers 9 distinct manipulation tasks. Both robot and human data are captured from an egocentric perspective. Specifically, hu… view at source ↗
Figure 5
Figure 5. Figure 5: Generalization Setup. including lighting, object position, and embodiment-related appearance shifts [PITH_FULL_IMAGE:figures/full_fig_p008_5.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

43 extracted references · 43 canonical work pages · 18 internal anchors

  1. [1]

    Yu et al

    T. Yu et al. One-shot imitation from observing humans via domain-adaptive meta-learning. In RSS, 2018

    work page 2018

  2. [2]

    S. Bahl, A. Gupta, and D. Pathak. Human-to-robot imitation in the wild. InRSS, 2022

    work page 2022

  3. [3]

    M. Xu, Z. Xu, C. Chi, M. Veloso, and S. Song. Xskill: Cross embodiment skill discovery. In CoRL, 2023

    work page 2023

  4. [4]

    Bharadhwaj, A

    H. Bharadhwaj, A. Gupta, V . Kumar, and S. Tulsiani. Towards generalizable zero-shot manip- ulation via translating human interaction plans. InICRA, 2024

    work page 2024

  5. [5]

    Hansen et al

    N. Hansen et al. Self-supervised policy adaptation during deployment. InICLR, 2021

    work page 2021

  6. [6]

    M. Xu, Z. Xu, C. C. Pan, X. Zhu, C. Tomei, Y . Shen, Z. Wu, S.-R. Chen, J. B. Tenenbaum, T. Lozano-Perez, and S. Song. Flow as the cross-domain manipulation interface. InCoRL, 2024

    work page 2024

  7. [7]

    EgoMimic: Scaling Imitation Learning via Egocentric Video

    S. Kareer, D. Patel, R. Punamiya, P. Mathur, S. Cheng, C. Wang, J. Hoffman, and D. Xu. Egomimic: Scaling imitation learning via egocentric video.arXiv preprint arXiv:2410.24221, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video

    R. Hoque, P. Huang, D. J. Yoon, M. Sivapurapu, and J. Zhang. Egodex: Learning dexterous manipulation from large-scale egocentric video.arXiv preprint arXiv:2505.11709, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Grauman et al

    K. Grauman et al. Ego-exo4d: Understanding skilled human activity from first- and third- person perspectives. InCVPR, 2024

    work page 2024

  10. [10]

    Egoscale: Scaling dexterous manipulation with diverse egocentric human data, 2026

    R. Zheng et al. Egoscale: Scaling dexterous manipulation with diverse egocentric human data. arXiv preprint arXiv:2602.16710, 2026

    work page arXiv 2026

  11. [11]

    Chen et al

    H. Chen et al. Vidbot: Learning generalizable 3d actions from in-the-wild 2d human videos for zero-shot robotic manipulation. InCVPR, 2025

    work page 2025

  12. [12]

    Kim et al

    H. Kim et al. Uniskill: Imitating human videos via cross-embodiment skill representations. In CoRL, 2025

    work page 2025

  13. [13]

    Z. Chen, S. Chen, E. Arlaud, I. Laptev, and C. Schmid. Vividex: Learning vision-based dex- terous manipulation from human videos. InICRA, 2025

    work page 2025

  14. [14]

    C. Wang, L. Fan, J. Sun, R. Zhang, L. Fei-Fei, D. Xu, Y . Zhu, and A. Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. InCoRL, 2023. arXiv:2302.12422

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    Zero-Shot Robot Manipulation from Passive Human Videos

    H. Bharadhwaj, A. Gupta, S. Tulsiani, and V . Kumar. Zero-shot robot manipulation from passive human videos.arXiv preprint arXiv:2302.02011, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    V . Jain, M. Attarian, N. J. Joshi, A. Wahid, D. Driess, Q. Vuong, P. R. Sanketi, P. Ser- manet, S. Welker, C. Chan, I. Gilitschenski, Y . Bisk, and D. Dwibedi. Vid2robot: End- to-end video-conditioned policy learning with cross-attention transformers.arXiv preprint arXiv:2403.12943, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    Gen2Act: Human Video Generation in Novel Scenarios enables Generalizable Robot Manipulation

    H. Bharadhwaj, D. Dwibedi, A. Gupta, S. Tulsiani, C. Doersch, T. Xiao, D. Shah, F. Xia, D. Sadigh, and S. Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation.arXiv preprint arXiv:2409.16283, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    R. Shah, S. Liu, Q. Wang, Z. Jiang, S. Kumar, M. Seo, R. Mart ´ın-Mart´ın, and Y . Zhu. Mim- icdroid: In-context learning for humanoid robot manipulation from human play videos.arXiv preprint arXiv:2509.09769, 2025. 10

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    EVA: An Embodied World Model for Future Video Anticipation

    X. Chi, C.-K. Fan, H. Zhang, X. Qi, R. Zhang, A. Chen, C.-m. Chan, W. Xue, Q. Liu, S. Zhang, et al. Eva: An embodied world model for future video anticipation.arXiv preprint arXiv:2410.15461, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    X. Chi, P. Jia, C.-K. Fan, X. Ju, W. Mi, K. Zhang, Z. Qin, W. Tian, K. Ge, H. Li, et al. Wow: Towards a world omniscient world model through embodied interaction.arXiv preprint arXiv:2509.22642, 2025

    work page arXiv 2025

  21. [21]

    C. Zhu, R. Yu, S. Feng, B. Burchfiel, P. Shah, and A. Gupta. Unified world models: Cou- pling video and action diffusion for pretraining on large robotic datasets.arXiv preprint arXiv:2504.02792, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    J. Lyu, K. Liu, X. Zhang, H. Liao, Y . Feng, W. Zhu, T. Shen, J. Chen, J. Zhang, Y . Dong, W. Cui, S. Qi, S. Wang, Y . Zheng, M. Yan, X. Shi, H. Li, D. Zhao, M.-Y . Liu, Z. Zhang, L. Yi, Y . Wang, and H. Wang. LDA-1B: Scaling latent dynamics action model via universal embodied data ingestion.arXiv preprint arXiv:2602.12215, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    H. Bi, H. Tan, S. Xie, Z. Wang, S. Huang, H. Liu, R. Zhao, Y . Feng, C. Xiang, Y . Rong, et al. Motus: A unified latent action world model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 35101–35113, 2026

    work page 2026

  24. [24]

    M. Team, C. Xiang, F. Bao, H. Liu, H. Tan, H. Bi, J. Li, J. Liu, J. Pang, K. Jing, et al. Mo- tubrain: An advanced world action model for robot control.arXiv preprint arXiv:2604.27792, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?

    Y . Zhang, W. Zhang, Z. Qi, H. Zhang, H. Lin, J. Zhang, Y . Mu, X. Yang, W. Zeng, and X. Jin. Imagewam: Do world action models really need video generation, or just image editing?, 2026. URLhttps://arxiv.org/abs/2606.19531

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  26. [26]

    H. Yu, H. Lin, J. Zhang, W. Zhang, C. Gu, H. Li, and P. Tan. Maskwam: Unifying mask prompting and prediction for world-action models.arXiv preprint arXiv:2606.13515, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  27. [27]

    B. Peng, W. Zhang, L. Xu, Z. Qi, J. Zhang, H. Liu, W. Zeng, and X. Jin. Reworld: Multi- dimensional reward modeling for embodied world models.arXiv preprint arXiv:2601.12428, 2026

    work page arXiv 2026

  28. [28]

    Y . Sun, X. Wang, Z. Liu, J. Miller, A. A. Efros, and M. Hardt. Test-time training with self- supervision for generalization under distribution shifts. InICML, 2020

    work page 2020

  29. [29]

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell. Tent: Fully test-time adaptation by entropy minimization. InICLR, 2021

    work page 2021

  30. [30]

    Y . Liu, P. Kothari, B. van Delft, B. Bellot-Gurlet, T. Mordan, and A. Alahi. Ttt++: When does self-supervised test-time training fail or thrive? InNeurIPS, 2021

    work page 2021

  31. [31]

    Gandelsman, Y

    Y . Gandelsman, Y . Sun, X. Chen, and A. A. Efros. Test-time training with masked autoen- coders. InNeurIPS, 2022

    work page 2022

  32. [32]

    Y . Sun, X. Li, K. Dalal, J. Xu, A. Vikram, G. Zhang, Y . Dubois, X. Chen, X. Wang, S. Koyejo, T. Hashimoto, and C. Guestrin. Learning to (learn at test time): Rnns with expressive hidden states. InICML, 2025

    work page 2025

  33. [33]

    Titans: Learning to Memorize at Test Time

    A. Behrouz, P. Zhong, and V . Mirrokni. Titans: Learning to memorize at test time.arXiv preprint arXiv:2501.00663, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    F. Liu, D. Wu, J. Chi, Y . Cai, Y .-H. Hung, X. Yu, H. Li, H. Hu, Y . Rao, and Y . Duan. Spatial-ttt: Streaming visual-based spatial intelligence with test-time training.arXiv preprint arXiv:2603.12255, 2026. 11

    work page arXiv 2026

  35. [35]

    S. Yang, Y . Ze, and H. Xu. Movie: Visual model-based policy adaptation for view generaliza- tion. InNeurIPS, 2023

    work page 2023

  36. [36]

    Z. Bai, C. Gao, and M. Z. Shou. Evolve-vla: Test-time training from environment feedback for vision-language-action models.arXiv preprint arXiv:2512.14666, 2025

    work page arXiv 2025

  37. [37]

    C. Liu, Y . Liu, T. Wang, Q. Zhuang, J. C. Liang, W. Yang, R. Xu, Q. Wang, D. Liu, and C. Han. On-the-fly vla adaptation via test-time reinforcement learning.arXiv preprint arXiv:2601.06748, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  38. [38]

    $\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Haus- man, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, L. Smith, J. T. Springenberg, K. Stachowicz, J. Tanner, Q. Vuong, H. Walke, A. Walling, H. Wang, L. Yu, and U. Zhilinsky.π0.5: a vision-language-action ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    J. Mu, S. Yang, Y . Bao, H. Bae, T. Wei, L. Xu, B. Li, H. Xu, and J. Pang. Deximit: Learning bimanual dexterous manipulation from monocular human videos.arXiv preprint arXiv:2602.10105, 2026

    work page arXiv 2026

  40. [40]

    Katharopoulos, A

    A. Katharopoulos, A. Vyas, N. Pappas, and F. Fleuret. Transformers are RNNs: Fast autore- gressive transformers with linear attention. InInternational Conference on Machine Learning (ICML), 2020

    work page 2020

  41. [41]

    MediaPipe: A Framework for Building Perception Pipelines

    C. Lugaresi, J. Tang, H. Nash, C. McClanahan, E. Uboweja, M. Hays, F. Zhang, C.-L. Chang, M. G. Yong, J. Lee, W.-T. Chang, W. Hua, M. Georg, and M. Grundmann. MediaPipe: A framework for building perception pipelines.arXiv preprint arXiv:1906.08172, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  42. [42]

    Romero, D

    J. Romero, D. Tzionas, and M. J. Black. Embodied hands: Modeling and capturing hands and bodies together. InACM Transactions on Graphics (SIGGRAPH Asia), 2017

    work page 2017

  43. [43]

    DINOv3

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haziza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H. J´egou, P. Labatut, and P. Bojanowski. DINOv3.arXiv preprint arXiv:2508.10104,...

    work page internal anchor Pith review Pith/arXiv arXiv 2025