pith. machine review for the scientific record. sign in

arxiv: 2605.01194 · v1 · submitted 2026-05-02 · 💻 cs.RO

Recognition: unknown

VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model

Wenhao Li , Xiu Su , Dan Niu , Yichao Cao , Hongyan Xu , Zhe Qu , Lei Fan , Shan You
show 1 more author
Chang Xu
Authors on Pith no claims yet

Pith reviewed 2026-05-09 15:06 UTC · model grok-4.3

classification 💻 cs.RO
keywords vision-language-actiontest-time computerelative action criticadaptive deliberationembodied manipulationuncertainty estimationpreference learningrobot failure reduction
0
0 comments X

The pith

Vision-language-action models can switch into deliberate test-time computation when uncertain, using pairwise action comparisons to reduce failures.

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

The paper aims to show that current vision-language-action models, which rely on rapid instinctive decisions, can be made more reliable in complex or ambiguous situations by dynamically allocating extra computation only when needed. It introduces an uncertainty detector to activate this deliberation phase, during which a relative critic evaluates candidate actions by comparing them in pairs rather than assigning absolute scores. This change simplifies training and avoids common instabilities, while an automated process generates the necessary comparison data and sampling keeps costs manageable. If the approach holds, embodied systems would fail less often on extended tasks without constantly expending high compute. Readers would care because it directly tackles the gap between fast but brittle robot policies and the need for thoughtful behavior in real environments.

Core claim

VLA-ATTC adds adaptive test-time compute to VLA models via an uncertainty-based cognitive clutch that shifts from reflexive action to a deliberation phase, where the Relative Action Critic selects the best action from candidates through pairwise comparisons, replacing unstable absolute scoring, supported by efficient sampling and an automated preference data pipeline, yielding over 50% failure rate reduction for the PI0.5 model on the LIBERO-LONG benchmark.

What carries the argument

The Relative Action Critic model, which identifies the optimal action among generated candidates by performing pairwise comparisons instead of direct value estimation.

If this is right

  • VLA models achieve substantially lower failure rates on long-horizon manipulation benchmarks.
  • The critic training objective becomes simpler and more stable than absolute value methods.
  • An automated pipeline can generate preference pairs for training without manual labeling.
  • Efficient sampling amortizes the added compute cost during the deliberation phase.
  • Existing state-of-the-art VLA models can be enhanced by this plug-in adaptive mechanism.

Where Pith is reading between the lines

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

  • The pairwise comparison idea could transfer to other generative decision systems that currently rely on scalar rewards.
  • Real-world robot deployments might see fewer catastrophic errors in ambiguous scenes if uncertainty triggers are tuned appropriately.
  • Scaling the approach to longer horizons or multi-agent settings would test whether the clutch and critic remain effective.
  • Combining this with other test-time methods like search or planning could compound reliability gains.

Load-bearing premise

Uncertainty estimates can reliably detect when extra deliberation is required and pairwise comparisons can stably identify superior actions without training collapse.

What would settle it

Running the full VLA-ATTC pipeline on the LIBERO-LONG benchmark and measuring whether the failure rate for PI0.5 drops by more than half; if the reduction is substantially smaller or the critic exhibits unstable training, the central claim would not hold.

Figures

Figures reproduced from arXiv: 2605.01194 by Chang Xu, Dan Niu, Hongyan Xu, Lei Fan, Shan You, Wenhao Li, Xiu Su, Yichao Cao, Zhe Qu.

Figure 1
Figure 1. Figure 1: Performance overview and efficiency comparison. impressive generalization capability across diverse complex manipulation tasks. However, their decision-making is gov￾erned by a fast, intuitive inference process (Xu et al., 2025). While this fast, intuitive strategy is sufficient for simple scenarios, it often leads to suboptimal or even catastrophic failures when facing complex or ambiguous situations that… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of VLA inference paradigms. (a) Standard Inference is fast but prone to failure in complex scenarios. (b) Sequential Deliberation needs costly retraining and data annotation. (c) Existing Parallel Deliberation incurs high costs and relies on unstable absolute scoring. (d) VLA-ATTC triggers deliberation only on hard situations, employing RAC model for robust candidate selection. to select the bes… view at source ↗
Figure 3
Figure 3. Figure 3: The overall architecture of the VLA-ATTC framework. The process begins with the “Cognitive Clutch” (Sec 4.1), which quantifies the VLA’s action generation uncertainty. If uncertainty is low , the model executes the initial action reflexively. If uncertainty exceeds the threshold, the Deliberation Phase (Sec 4.2) is activated. This phase 1) efficiently generates N action candidates in parallel by amortizing… view at source ↗
Figure 4
Figure 4. Figure 4: Detailed architecture of the Relative Action Critic (RAC) model (Sec 4.3). (Left) The VLM backbone is augmented with Nq learnable queries. During the single pre-filling pass, these queries distill high-level semantic context query features from the image and instruction, alongside the standard raw features. (Middle) The RAC model itself is a Transformer with consistent depth of VLM. It takes four inputs pr… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of success rate over different RAC model’s variant on LIBERO-LONG. (94% for PI0.5+ATTC and 90.8% for PI0+ATTC), removing the “Learnable Query” component causes the most signifi￾cant performance drop, lowering success rates to 92.4% and 88% respectively. These findings confirm that each design choice in the RAC framework is necessary to effectively evaluate and select optimal actions. 5.3 Mechani… view at source ↗
Figure 6
Figure 6. Figure 6: Two representative cases of PI0.5-ATTC. It adaptively triggers TTC when facing difficult situations view at source ↗
read the original abstract

Vision-Language-Action (VLA) models have demonstrated remarkable capabilities and generalization in embodied manipulation. However, their decision-making relies on a fast, instinctive process that lacks deliberation. This strategy often leads to suboptimal or catastrophic actions when facing complex or ambiguous scenarios that require greater consideration. In this paper, we introduce \textbf{VLA-ATTC}, a framework that endows VLA models with adaptive test-time compute (TTC). VLA-ATTC employs an uncertainty-based ``cognitive clutch'' to dynamically transition from reflexive execution to a TTC deliberation phase when necessary. During TTC phase, a novel \textbf{Relative Action Critic} (RAC) model identifies the optimal action from generated candidates via pairwise comparisons. This relative mechanism replaces unstable absolute value estimation, significantly simplifying the learning objective. Furthermore, we introduce an efficient sampling strategy to amortize computational costs and an automated data pipeline that curates preference pairs without manual annotation. On the LIBERO-LONG benchmark, VLA-ATTC reduces the failure rate of the SOTA model PI0.5 by over 50\%. We will open-source all the code and weights.

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 / 2 minor

Summary. The paper introduces VLA-ATTC, a framework for endowing Vision-Language-Action (VLA) models with adaptive test-time compute. It uses an uncertainty-based 'cognitive clutch' to dynamically switch from reflexive action execution to a deliberation phase when facing complex or ambiguous states. In the deliberation phase, a novel Relative Action Critic (RAC) model selects the optimal action from candidate actions via pairwise comparisons, replacing absolute value estimation. The approach includes an efficient sampling strategy to amortize costs and an automated pipeline for generating preference pairs without manual annotation. The central empirical claim is that VLA-ATTC reduces the failure rate of the SOTA PI0.5 model by over 50% on the LIBERO-LONG benchmark.

Significance. If the reported results hold under rigorous validation, the work could meaningfully advance reliable deployment of VLA models in embodied manipulation by enabling selective deliberation without constant high compute overhead. The relative pairwise formulation of the critic simplifies the learning objective relative to absolute critics and the automated preference data pipeline removes a common annotation bottleneck. The commitment to open-sourcing code and weights would support reproducibility and follow-on work.

major comments (3)
  1. [Abstract] Abstract: The headline claim of >50% failure-rate reduction on LIBERO-LONG is load-bearing for the paper's contribution, yet the abstract (and by extension the presented description) provides no quantitative validation that the uncertainty-based clutch reliably identifies only the states where extra compute improves outcomes (e.g., no precision/recall or failure-case analysis of clutch triggers).
  2. [Abstract] Abstract: The RAC is asserted to 'significantly simplify the learning objective' and avoid instability of absolute critics, but no analysis, ablation, or metric (such as transitivity error rate or ranking consistency across pairwise comparisons) is supplied to confirm that the pairwise preference model converges to stable, transitive orderings on the generated candidates.
  3. [Abstract] Abstract: The automated data pipeline for curating preference pairs is presented as a key enabler, yet no details are given on how negative or ambiguous pairs are filtered, how label noise is controlled, or whether the resulting training distribution matches the distribution of states encountered at test time under the clutch.
minor comments (2)
  1. [Abstract] Abstract: The phrase 'cognitive clutch' is evocative but should be accompanied by a concise technical definition of the uncertainty signal (e.g., entropy, variance, or model disagreement) in the main text.
  2. [Abstract] Abstract: Reporting only the aggregate failure-rate reduction leaves open whether gains are uniform across task categories or driven by a few long-horizon episodes; per-task breakdowns or confidence intervals would strengthen the result.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below and outline the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline claim of >50% failure-rate reduction on LIBERO-LONG is load-bearing for the paper's contribution, yet the abstract (and by extension the presented description) provides no quantitative validation that the uncertainty-based clutch reliably identifies only the states where extra compute improves outcomes (e.g., no precision/recall or failure-case analysis of clutch triggers).

    Authors: We agree that the abstract would benefit from additional context on the clutch. The full manuscript (Section 4.2) includes quantitative analysis and failure-case studies showing that the uncertainty clutch triggers deliberation primarily on states where the base VLA model fails, directly contributing to the reported performance gains. We will revise the abstract to include a brief statement summarizing the clutch's empirical reliability and its selective activation. revision: yes

  2. Referee: [Abstract] Abstract: The RAC is asserted to 'significantly simplify the learning objective' and avoid instability of absolute critics, but no analysis, ablation, or metric (such as transitivity error rate or ranking consistency across pairwise comparisons) is supplied to confirm that the pairwise preference model converges to stable, transitive orderings on the generated candidates.

    Authors: The manuscript motivates the relative formulation as a simplification that avoids the calibration challenges of absolute critics. We acknowledge that explicit supporting metrics would strengthen this claim. We will add an ablation study and analysis of transitivity error and ranking consistency to the revised manuscript to empirically validate the stability of the RAC model on generated candidates. revision: yes

  3. Referee: [Abstract] Abstract: The automated data pipeline for curating preference pairs is presented as a key enabler, yet no details are given on how negative or ambiguous pairs are filtered, how label noise is controlled, or whether the resulting training distribution matches the distribution of states encountered at test time under the clutch.

    Authors: The automated pipeline is detailed in Section 3.4 of the manuscript, where pairs are derived from base-model rollouts. To directly address the concern, we will expand both the relevant section and the abstract to specify the filtering criteria for negative/ambiguous pairs, noise-control measures, and evidence of distribution alignment between training pairs and clutch-triggered test-time states. revision: yes

Circularity Check

0 steps flagged

No circularity: new algorithmic components with empirical validation

full rationale

The paper introduces an uncertainty-based cognitive clutch and a Relative Action Critic (RAC) that uses pairwise comparisons instead of absolute value estimation. These are presented as novel engineering choices with an automated data pipeline for preference pairs. No equations, derivations, or central claims reduce by construction to the inputs (e.g., no fitted parameter renamed as a prediction, no self-citation chain invoked as a uniqueness theorem, no ansatz smuggled via prior work). Performance gains on LIBERO-LONG are reported as empirical results, not tautological consequences of the method definition. The derivation chain is self-contained and externally falsifiable via benchmark evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Based on abstract, the method introduces new components whose effectiveness relies on empirical results not detailed here.

axioms (1)
  • domain assumption VLA models can benefit from additional compute in uncertain scenarios
    Assumed in the framework design
invented entities (1)
  • Relative Action Critic model no independent evidence
    purpose: To identify optimal action via pairwise comparisons
    New model introduced without external validation mentioned

pith-pipeline@v0.9.0 · 5520 in / 1065 out tokens · 26334 ms · 2026-05-09T15:06:43.310658+00:00 · methodology

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

32 extracted references · 31 canonical work pages · 9 internal anchors

  1. [1]

    $\pi_0$: A Vision-Language-Action Flow Model for General Robot Control

    Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., et al. pi0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    work page internal anchor Pith review arXiv

  2. [2]

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

    Black, K., Brown, N., Darpinian, J., Dhabalia, K., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Galliker, M. Y ., et al. pi0.5: a vision-language-action model with open- world generalization.arXiv preprint arXiv:2504.16054,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models

    Chen, Q., Qin, L., Liu, J., Peng, D., Guan, J., Wang, P., Hu, M., Zhou, Y ., Gao, T., and Che, W. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567,

    work page internal anchor Pith review arXiv

  4. [4]

    Multi-agent embodied ai: Advances and future directions.arXiv preprint arXiv:2505.05108,

    Feng, Z., Xue, R., Yuan, L., Yu, Y ., Ding, N., Liu, M., Gao, B., Sun, J., Zheng, X., and Wang, G. Multi-agent embodied ai: Advances and future directions.arXiv preprint arXiv:2505.05108,

    work page arXiv

  5. [5]

    arXiv preprint arXiv:2506.22355 (2025) 5

    Fung, P., Bachrach, Y ., Celikyilmaz, A., Chaudhuri, K., Chen, D., Chung, W., Dupoux, E., Gong, H., J´egou, H., Lazaric, A., et al. Embodied ai agents: Modeling the world.arXiv preprint arXiv:2506.22355,

    work page arXiv

  6. [6]

    Vla-reasoner: Empowering vision-language-action models with reasoning via online monte carlo tree search.arXiv preprint arXiv:2509.22643, 2025

    Guo, W., Lu, G., Deng, H., Wu, Z., Tang, Y ., and Wang, Z. Vla-reasoner: Empowering vision-language-action models with reasoning via online monte carlo tree search. arXiv preprint arXiv:2509.22643,

    work page arXiv

  7. [7]

    Efficient test-time scaling via self-calibration

    Huang, C., Huang, L., Leng, J., Liu, J., and Huang, J. Effi- cient test-time scaling via self-calibration.arXiv preprint arXiv:2503.00031,

    work page arXiv

  8. [8]

    OpenVLA: An Open-Source Vision-Language-Action Model

    Kim, M. J., Pertsch, K., Karamcheti, S., Xiao, T., Balakr- ishna, A., Nair, S., Rafailov, R., Foster, E., Lam, G., San- keti, P., et al. Openvla: An open-source vision-language- action model.arXiv preprint arXiv:2406.09246,

    work page internal anchor Pith review arXiv

  9. [9]

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

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

    work page internal anchor Pith review arXiv

  10. [10]

    Robomonkey: Scaling test-time sampling and ver- ification for vision-language-action models.arXiv preprint arXiv:2506.17811, 2025

    Kwok, J., Agia, C., Sinha, R., Foutter, M., Li, S., Stoica, I., Mirhoseini, A., and Pavone, M. Robomonkey: Scaling test-time sampling and verification for vision-language- action models.arXiv preprint arXiv:2506.17811,

    work page arXiv

  11. [11]

    Gonzalez, and Ion Stoica

    Li, D., Cao, S., Cao, C., Li, X., Tan, S., Keutzer, K., Xing, J., Gonzalez, J. E., and Stoica, I. S*: Test time scaling for code generation.arXiv preprint arXiv:2502.14382,

    work page arXiv

  12. [12]

    CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation

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

    work page Pith review arXiv

  13. [13]

    Onetwovla: A unified vision-language-action model with adaptive reasoning

    Lin, F., Nai, R., Hu, Y ., You, J., Zhao, J., and Gao, Y . Onetwovla: A unified vision-language-action model with adaptive reasoning.arXiv preprint arXiv:2505.11917,

    work page arXiv

  14. [14]

    s1: Simple test-time scaling

    Muennighoff, N., Yang, Z., Shi, W., Li, X. L., Fei-Fei, L., Hajishirzi, H., Zettlemoyer, L., Liang, P., Cand `es, E., and Hashimoto, T. s1: Simple test-time scaling.arXiv preprint arXiv:2501.19393,

    work page Pith review arXiv

  15. [15]

    Steering your generalists: Improving robotic foundation models via value guidance.arXiv preprint arXiv:2410.13816, 2024

    Nakamoto, M., Mees, O., Kumar, A., and Levine, S. Steer- ing your generalists: Improving robotic foundation mod- els via value guidance.arXiv preprint arXiv:2410.13816,

    work page arXiv

  16. [16]

    T., Nguyen, B., and Nguyen, V

    Nguyen, H. T., Nguyen, B., and Nguyen, V . A. Struc- tured pruning for diverse best-of-n reasoning optimiza- tion.arXiv preprint arXiv:2506.03978,

    work page arXiv

  17. [17]

    Y ., Xu, J., Fazel-Zarandi, M., Bansal, M., Sukhbaatar, S., Weston, J., and Yu, J

    Prasad, A., Yuan, W., Pang, R. Y ., Xu, J., Fazel-Zarandi, M., Bansal, M., Sukhbaatar, S., Weston, J., and Yu, J. Self-consistency preference optimization.arXiv preprint arXiv:2411.04109,

    work page arXiv

  18. [18]

    G., Dadashi, R., Hussenot, L., Ferret, J., Vieil- lard, N., Ram ´e, A., Shariari, B., Perrin, S., Friesen, A., Cideron, G., et al

    Sessa, P. G., Dadashi, R., Hussenot, L., Ferret, J., Vieil- lard, N., Ram ´e, A., Shariari, B., Perrin, S., Friesen, A., Cideron, G., et al. Bond: Aligning llms with best-of-n distillation.arXiv preprint arXiv:2407.14622,

    work page arXiv

  19. [19]

    SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics

    Shukor, M., Aubakirova, D., Capuano, F., Kooijmans, P., Palma, S., Zouitine, A., Aractingi, M., Pascal, C., Russi, M., Marafioti, A., et al. Smolvla: A vision-language- action model for affordable and efficient robotics.arXiv preprint arXiv:2506.01844,

    work page internal anchor Pith review arXiv

  20. [20]

    Hume: Introducing system-2 thinking in visual-language-action model.arXiv preprint arXiv:2505.21432, 2025

    Song, H., Qu, D., Yao, Y ., Chen, Q., Lv, Q., Tang, Y ., Shi, M., Ren, G., Yao, M., Zhao, B., et al. Hume: Introducing system-2 thinking in visual-language-action model.arXiv preprint arXiv:2505.21432,

    work page arXiv

  21. [21]

    Octo: An Open-Source Generalist Robot Policy

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

    work page internal anchor Pith review arXiv

  22. [22]

    Genselect: A generative approach to best-of-n

    Toshniwal, S., Sorokin, I., Ficek, A., Moshkov, I., and Gitman, I. Genselect: A generative approach to best-of-n. arXiv preprint arXiv:2507.17797,

    work page arXiv

  23. [23]

    M., Oesterling, A., Lakkaraju, H., and Calmon, F

    Verdun, C. M., Oesterling, A., Lakkaraju, H., and Calmon, F. P. Soft best-of-n sampling for model alignment.arXiv preprint arXiv:2505.03156,

    work page arXiv

  24. [24]

    Soft self-consistency improves language model agents

    Wang, H., Prasad, A., Stengel-Eskin, E., and Bansal, M. Soft self-consistency improves language model agents. arXiv preprint arXiv:2402.13212, 2024a. Wang, X. and Zhou, D. Chain-of-thought reasoning without prompting.Advances in Neural Information Processing Systems, 37:66383–66409,

    work page arXiv

  25. [25]

    Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning

    Wang, X., Feng, S., Li, Y ., Yuan, P., Zhang, Y ., Tan, C., Pan, B., Hu, Y ., and Li, K. Make every penny count: Difficulty-adaptive self-consistency for cost-efficient rea- soning.arXiv preprint arXiv:2408.13457, 2024b. Wang, X., Feng, S., Li, Y ., Yuan, P., Zhang, Y ., Tan, C., Pan, B., Hu, Y ., and Li, K. Make every penny count: Difficulty-adaptive sel...

    work page arXiv 2025

  26. [26]

    Vla-cache: Efficient vision-language-action manipulation via adaptive token caching,

    Xu, S., Wang, Y ., Xia, C., Zhu, D., Huang, T., and Xu, C. Vla-cache: Towards efficient vision-language-action model via adaptive token caching in robotic manipulation. arXiv preprint arXiv:2502.02175,

    work page arXiv

  27. [27]

    Demystifying long chain-of-thought reasoning in llms, 2025

    Yeo, E., Tong, Y ., Niu, M., Neubig, G., and Yue, X. Demys- tifying long chain-of-thought reasoning in llms.arXiv preprint arXiv:2502.03373,

    work page arXiv

  28. [28]

    Robotic Control via Embodied Chain-of-Thought Reasoning

    Zawalski, M., Chen, W., Pertsch, K., Mees, O., Finn, C., and Levine, S. Robotic control via embodied chain-of-thought reasoning.arXiv preprint arXiv:2407.08693,

    work page internal anchor Pith review arXiv

  29. [29]

    Revisit- ing the test-time scaling of o1-like models: Do they truly possess test-time scaling capabilities?arXiv preprint arXiv:2502.12215,

    Zeng, Z., Cheng, Q., Yin, Z., Zhou, Y ., and Qiu, X. Revisit- ing the test-time scaling of o1-like models: Do they truly possess test-time scaling capabilities?arXiv preprint arXiv:2502.12215,

    work page arXiv

  30. [30]

    A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?

    Zhang, Q., Lyu, F., Sun, Z., Wang, L., Zhang, W., Hua, W., Wu, H., Guo, Z., Wang, Y ., Muennighoff, N., et al. A survey on test-time scaling in large language mod- els: What, how, where, and how well?arXiv preprint arXiv:2503.24235,

    work page internal anchor Pith review arXiv

  31. [31]

    Accelerating best-of-n via specula- tive rejection

    Zhang, R., Haider, M., Yin, M., Qiu, J., Wang, M., Bartlett, P., and Zanette, A. Accelerating best-of-n via specula- tive rejection. InICML 2024 Workshop on Structured Probabilistic Inference Generative Modeling,

    2024

  32. [32]

    Vision-language-action model with open- world embodied reasoning from pretrained knowledge

    Zhou, Z., Zhu, Y ., Wen, J., Shen, C., and Xu, Y . Chatvla-2: Vision-language-action model with open-world embodied reasoning from pretrained knowledge.arXiv preprint arXiv:2505.21906, 2025a. Zhou, Z., Zhu, Y ., Zhu, M., Wen, J., Liu, N., Xu, Z., Meng, W., Cheng, R., Peng, Y ., Shen, C., et al. Chatvla: Unified 10 VLA-ATTC: Adaptive Test-Time Compute for ...

    work page arXiv