Pith. sign in

REVIEW 4 major objections 5 minor 38 references

This paper claims that the dynamic importance of vision in frozen VLA models can be estimated online via zero-padding counterfactuals, and that using this causal signal to gate a residual correction improves robot manipulation success witho

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 · deepseek-v4-flash

2026-08-01 02:10 UTC pith:RKS3IJGL

load-bearing objection A useful test-time vision booster for frozen VLA policies, wrapped in causal language that goes beyond what zero-padding can support; worth a serious referee, not a desk reject. the 4 major comments →

arxiv 2607.25516 v1 pith:RKS3IJGL submitted 2026-07-28 cs.RO

A Causality-aware Infer-diagnose-refine Framework for Test-time Modality Adaptation in VLA Models

Haoyu Zhang , Yuwei Wu , Jin Chen , Gao Zhi , Zhenxin Diao , Mingyang Gao , Kun Wu , Yongchun Liu
show 1 more author
Fan Li
This is my paper
classification cs.RO
keywords Vision-Language-Action modelstest-time adaptationcausal inferencecounterfactual interventionmodality fusionrobot manipulationgated residual fusionzero-padding
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 tries to establish that visual importance in a vision-language-action (VLA) model is not fixed but varies over time, across architectures, and across environments—and that this importance can be measured at test time by asking what the model would predict if the visual input were absent. It proposes IDR, a training-free pipeline that infers actions under factual and counterfactual inputs, diagnoses the L2 norm of the action deviation as a per-timestep causal effect, and refines the action through gated residual fusion. The authors report consistent gains across four frozen VLA backbones on LIBERO, SIMPLER, CALVIN, and real-world tasks, including a real-world success increase from 56.5% to 75.3%. A sympathetic reader would care because, if correct, modality fusion in VLA models becomes an adjustable test-time mechanism rather than a fixed training-time design.

Core claim

The central claim is that visual importance is a time-varying, environment-dependent, and architecture-dependent quantity that can be diagnosed at each timestep as the causal effect of the visual observation on the predicted action. The paper operationalizes this by zero-padding the visual input (and separately the proprioceptive state) to create counterfactual scenarios, then computing the L2 norm of the difference between the factual action and each counterfactual action. When the visual causal effect falls below a threshold, a gate opens and injects a bounded residual correction scaled by the diagnosed effect. On frozen models—pi0.5, X-VLA, VLA-Adapter, and OpenVLA-OFT—IDR reports average

What carries the argument

The load-bearing object is the visual causal effect E_img,t = ||a_base,t - a_no_img,t||_2, the L2 norm of the deviation between the factual action and the action produced when the visual observation is replaced by an all-zero tensor. Its proprioceptive analog E_prop,t is computed the same way. Together they drive a gated residual fusion: when E_img,t falls below a threshold tau, the gate opens and the refined action is a_final = a_base + g_t * (alpha * Delta_img + w_prop * clip(Delta_prop, -lambda, lambda)), where w_prop is bounded by E_prop/(E_img + epsilon). This mechanism converts 'visual importance' into an observable, per-timestep signal that can correct blind execution while preserving

Load-bearing premise

The central assumption is that zero-padding the visual input is a faithful counterfactual for 'no visual observation' in the frozen VLA model, so that the norm of the resulting action deviation measures visual causal importance rather than an artifact of feeding an out-of-distribution input.

What would settle it

Run IDR alongside a control intervention that corrupts the image with task-irrelevant noise of matching L2 norm (e.g., random Gaussian noise). If the control intervention triggers the gate as often as zero-padding and produces the same success gains, then the improvement does not depend on the counterfactual being a faithful absence of vision; if the true visual ablation changes E_img and downstream success while the control does not, the causal story is supported.

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

If this is right

  • Frozen VLA models can be improved at test time without retraining by dynamically adjusting visual importance.
  • Visual importance is structured but not fixed: it shifts across architectures, environments, and task phases such as gripper open/close events.
  • Selective gating (refine only when diagnosed visual effect is low) outperforms uniform correction; applying the correction at every timestep degrades performance below baseline.
  • The direction of the causal correction is critical: reversing the sign of the injected effect reduces success below the unmodified baseline.
  • In real-world deployment, three forward passes per step can still reduce total completion time because refined actions are more task-directed and reduce unnecessary motions.

Where Pith is reading between the lines

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

  • Because IDR requires no labels or retraining, it could serve as a general probe for modality reliance in any multimodal policy, not just VLA models; the same norm-based effect could audit when a model ignores language, depth, or tactile input.
  • The zero-padding counterfactual is likely to be most faithful for models trained with masking or normalization; a direct extension would test whether replacing zero-padding with token dropout or a learned 'no-vision' token preserves the gate's behavior.
  • The intervention threshold is set to the baseline model's mean visual effect, which introduces a degree of circularity; a per-task calibrated threshold or a continuous gate instead of a hard indicator could make the method more robust and reveal how sensitive the gains are to this choice.
  • The authors explicitly flag the three-forward-pass latency as the primary limitation and suggest distilling the online causal diagnostic into the model during training; if that distillation succeeds, it would imply the gating signal can be learned implicitly without test-time overhead.

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

4 major / 5 minor

Summary. The paper proposes IDR, a training-free, model-agnostic test-time adaption framework for frozen vision-language-action (VLA) models. At each timestep, IDR computes the factual action and two counterfactual actions obtained by zero-padding the visual and proprioceptive inputs (Eq. 7), measures the L2 norms of the resulting deviations as modality-specific causal effects (Eqs. 8-9), and uses a gated residual fusion to refine the action when the visual effect is below a threshold (Eqs. 10-13). The authors report average success-rate improvements across four VLA backbones on LIBERO, SIMPLER, CALVIN, and real-world manipulation tasks, and include ablations of intervention strategies, hyperparameters, and components.

Significance. If validated, the paper would provide a practical plug-and-play method for improving frozen VLA policies without retraining, which is relevant to robotics and multimodal policy learning. The framework is simple, clearly described, and the evaluation spans multiple backbones, simulation benchmarks, and real-world tasks. The ablations are useful, and the real-world results, though small-scale, are encouraging. However, the central 'causality-aware' claim depends on treating zero-padding as a valid counterfactual intervention, which is not established; moreover, several reported gains are small and lack statistical support, and key hyperparameters are tuned on the evaluation benchmarks.

major comments (4)
  1. [§IV.B.1, Eq. (7)] The causal claim rests on treating the all-zero visual input 0_v as the counterfactual 'no visual observation.' For a frozen VLA model, an all-zero image is an out-of-distribution input, not a missing observation, so pi_theta*(0_v, s, l) may be an arbitrary extrapolation. Consequently E_img (Eq. 9) and the gate g_t (Eq. 10) may reflect sensitivity to a specific OOD artifact rather than the causal effect of visual content. Table VI does not resolve this: Gaussian-noise and mean-value interventions also improve the baseline, so the empirical benefit is compatible with non-causal explanations. The authors should validate zero-padding against a more principled absence intervention (e.g., masking visual tokens in the internal representation, or an input intervention that stays on the training manifold) and show that the resulting diagnostic tracks task-relevant visual reliance.
  2. [§V.C, Fig. 8; Eq. (10)] Hyperparameters alpha and tau are tuned on the same benchmarks used for final evaluation. In particular, tau is defined as the mean visual causal effect of the baseline model and the ablation in Fig. 8 chooses the best value on LIBERO with pi0.5; the gains reported in Table II may therefore be optimistic. The authors should specify how tau is computed for each backbone/environment (e.g., on a held-out validation set), report results for fixed alpha/tau across all benchmarks, or perform a nested evaluation with a separate validation split.
  3. [Tables II, III, IV, V] Several headline improvements are within typical run-to-run noise and no error bars, confidence intervals, or number of evaluation seeds are reported. For example, X-VLA on LIBERO gains only +0.45 on the average (with Goal dropping 98.0->95.0), Google Visual Matching gains +0.98, and CALVIN VLA-Adapter gains +0.04 in average length. The real-world Organize Table result is based on 6 trials, making the 0->33.3% full-completion claim fragile. The claim of 'reliable improvements across multiple backbones' requires repeated evaluations, standard deviations, and significance testing.
  4. [§IV.B.3, Eqs. (8)-(13)] There is a potential logical tension in the refinement mechanism: the gate triggers when E_img is below tau, i.e., when ||Delta_img|| is small, yet the injected visual correction is alpha*Delta_img, whose magnitude is also small in exactly those cases. The textual explanation that the correction 'compensates for suppressed visual importance' is therefore not directly supported by the equations. The authors should provide quantitative evidence on the actual correction magnitudes relative to action scales, and an analysis (e.g., in the style of Fig. 5) showing that the gate activates with nonzero, task-relevant Delta_img rather than adding a small perturbation that merely happens to help on average.
minor comments (5)
  1. [Eq. (7)] The do(.) notation is overloaded: do(V=0_v) is used for a tensor-valued variable, but the causal graph in Fig. 3 treats V as a single node. Clarify that the intervention replaces the entire visual observation tensor with zeros.
  2. [Table I] R_img is used without a formal definition in the table caption. Define it and the epsilon used for numerical stability.
  3. [Eq. (11)-(12)] The action scale is not reported. Since alpha=0.08, beta=0.05, and lambda=0.1 are absolute action-space constants, their appropriateness depends on the action normalization; please report action ranges and units.
  4. [§V.B] The real-world baseline uses Qwen3-VL-4B with a Transformer action head, which is not one of the four backbones evaluated in simulation. Clarify whether this is a separate architecture and how its action outputs compare with those in simulation.
  5. [General] The paper would benefit from stating the number of evaluation episodes for LIBERO, SIMPLER, and CALVIN, and from indicating whether the same evaluation protocol (e.g., number of rollouts per task) was used for all baselines.

Circularity Check

1 steps flagged

Diagnosis and refinement are two uses of the same counterfactual deviation; the causal framing is definitionally coupled, though the empirical gain is measured rather than derived.

specific steps
  1. self definitional [Section IV-B, Eqs. (8)–(13)]
    "∆img,t = a base,t − a no img,t ... Eimg,t = ∥∆img,t∥2 ... gt = I[Eimg,t < τ] ... afinal,t = a base,t + gt ·(α∆img,t + w prop,t ·u prop,t)"

    The diagnostic signal E_img,t is defined as the L2 norm of the deviation vector Δ_img,t, and the refinement step adds a scaled version of the same vector Δ_img,t when the gate activates. Hence the 'diagnosed causal effect' and the 'targeted visual refinement' are not independent quantities: the gate trigger is a scalar function of the exact vector used as the correction. The statement that refinement is guided by an independently estimated visual importance is therefore true by construction, so the causal interpretation is not separately evidenced by the mechanism itself. This is a definitional coupling rather than a full prediction-from-input reduction; the empirical success is still measured against baselines, but the causal claim rests on the same quantity used for correction.

full rationale

I walked the derivation chain: Eq. (6) defines the base action; Eq. (7) constructs zero-padding counterfactuals; Eqs. (8)–(9) define the deviation vectors and the scalar norms; Eq. (10) gates on the norm; Eq. (13) uses the deviation vector as the corrective signal. The only apparent reduction is that the diagnostic magnitude E_img,t is the norm of the correction vector Δ_img,t, so 'diagnose' and 'refine' are two views of the same counterfactual comparison. This is a conceptual coupling but not a mathematically forced success: the reported improvements are empirical and could have failed. The zero-padding OOD concern (all-zero images are not a true 'no observation' intervention for a frozen VLA) is a validity threat, not a circularity. The threshold τ is set to the mean visual causal effect on the evaluation distribution, which is a test-distribution-dependent hyperparameter choice, but it is not a fitted value renamed as a prediction. No load-bearing self-citation appears: reference [20] is a related-work mention, not an argument that forces the result. Overall, the central derivation is not circular by construction; the score reflects the definitional coupling of diagnosis and refinement and the threshold selection on the evaluation benchmark.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The central claim rests on two domain assumptions and several fitted constants. The zero-padding intervention (Eq. 7) is assumed to realize a 'no visual observation' counterfactual; the L2 norm of the action deviation (Eq. 9) is assumed to be proportional to visual importance; the gate threshold τ (Eq. 10) is set from the baseline model's mean effect on the evaluation distribution. Free parameters α, β, λ, τ are fixed from ablations on the same benchmarks where success is reported.

free parameters (4)
  • τ (intervention threshold) = ≈7 for π0.5 LIBERO; defined as mean baseline E_img
    Controls gate activation; ablation in Fig. 8 shows best at τ=7, so it is tuned per benchmark rather than derived.
  • α (visual refinement strength) = 0.08
    Set from ablation sweep in Fig. 8; large values degrade performance.
  • β (proprioceptive regularization strength) = 0.05
    Global constant, no robustness analysis provided.
  • λ (proprioceptive residual clip) = 0.1
    Global constant that bounds Δprop,t to [−λ, λ].
axioms (3)
  • domain assumption Zero-padding the visual input (do(V=0_v)) realizes the counterfactual 'no visual observation' for a frozen VLA model.
    Used in Eq. (7); if zero-padded images are out-of-distribution and produce arbitrary activations, E_img (Eq. 9) is not a causal effect.
  • domain assumption The L2 norm of the action deviation ∥a_base − a_no_img∥ measures the causal importance of vision at timestep t.
    Used in Eq. (9); assumes action-space sensitivity is a faithful proxy for modality reliance.
  • ad hoc to paper The baseline model's mean visual causal effect is a valid intervention threshold τ for deciding when visual cues are underused.
    Eq. (10); validated only by ablation on the same benchmark where success is reported, so it is not an independently grounded rule.

pith-pipeline@v1.3.0-alltime-deepseek · 12399 in / 12553 out tokens · 123154 ms · 2026-08-01T02:10:57.574703+00:00 · methodology

0 comments
read the original abstract

Vision-language-action (VLA) models predict sequential actions to execute tasks specified by language instructions, conditioned on visual observations and proprioceptive states. However, how to fuse modalities in VLA models remains an open problem, since robot manipulation involves dynamic phases, such as long-distance movements and close-range interactions, in which the importance of visual observations may vary over time. In this paper, we propose an infer-diagnose-refine (IDR) framework, a model-agnostic framework that can be integrated with diverse VLA architectures for refining action predictions at test time. IDR first infers actions under factual and counterfactual scenarios of visual observations, and then diagnoses the causal effects of visual observations as the estimated dynamic importance, which is finally used to refine the action predictions in a training-free manner. We further design a causality-aware action refiner to realize the IDR framework, including zero-padding interventions for inferring counterfactual actions, norm-based quantification for diagnosing causal effects, and gated residual fusion for refining actions. Extensive experiments on both simulation benchmarks and real-world tasks show improvements in overall performance across multiple VLA backbones, demonstrating the efficacy of dynamically adjusting visual importance at test time.

Figures

Figures reproduced from arXiv: 2607.25516 by Fan Li, Gao Zhi, Haoyu Zhang, Jin Chen, Kun Wu, Mingyang Gao, Yongchun Liu, Yuwei Wu, Zhenxin Diao.

Figure 1
Figure 1. Figure 1: Existing data-driven VLA models directly produce a base action [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of IDR framework. IDR keeps the VLA model frozen and refines its base action with causal effects estimated at test time. The diagnosed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Causal graph of IDR. Causal effects are estimated by comparing [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Rimg heatmap across VLA backbones and LIBERO suites. Each model maintains a distinct visual importance pattern. TABLE I X-VLA VISUAL IMPORTANCE PATTERNS ACROSS ENVIRONMENTS. Environment Suite/Task Eimg Eprop Rimg LIBERO All suites 0.67 1.59 29.6% SIMPLER WidowX 0.62 0.76 44.9% CALVIN ABC→D 0.61 0.60 50.4% SIMPLER Google Variant Aggregation 1.18 0.09 92.9% SIMPLER Google Visual Matching 0.99 0.07 93.4% Eimg… view at source ↗
Figure 5
Figure 5. Figure 5: Phase-aligned visual importance ratio Rimg of π0.5 in LIBERO. The plot reports Rimg over task progress for the baseline, Mode E, and Mode F. Dashed vertical lines indicate gripper closing and opening events. importance patterns. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Real-world qualitative comparisons on Fold Clothes and Organize Table. The baseline model stalls before task completion, leading to failed [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: analyzes the two primary hyperparameters. For the correction scale α, performance peaks at 0.08, while nearby values such as 0.05 and 0.10 produce comparable results. Large values (α ≥ 0.5) degrade performance, indicating that excessive correction can destabilize action generation. Over￾all, the model remains robust within a moderate range of α. For the intervention threshold, small values rarely activate … 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

38 extracted references · 11 linked inside Pith

  1. [1]

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

    K. Black, N. Brown, D. Driess,et al., “π 0: A vision-language- action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Openvla: An open- source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti,et al., “Openvla: An open- source vision-language-action model,” inProc. Conf. Robot Learn., 2025

    2025

  3. [3]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu,et al., “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” inProc. Conf. Robot Learn., 2023, pp. 2165–2183

    2023

  4. [4]

    Octo: An open-source generalist robot policy,

    Octo Model Team, D. Ghosh, H. Walke,et al., “Octo: An open-source generalist robot policy,” inProc. Robot.: Sci. Syst., 2024

    2024

  5. [5]

    π 0.5: A vision- language-action model with open-world generalization,

    Physical Intelligence, K. Black, N. Brown,et al., “π 0.5: A vision- language-action model with open-world generalization,” inProc. Conf. Robot Learn., 2025, pp. 17–40

    2025

  6. [6]

    X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model,

    J. Zheng, J. Li, Z. Wang,et al., “X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model,” inProc. Int. Conf. Learn. Represent., 2026

    2026

  7. [7]

    Vla-adapter: An effective paradigm for tiny-scale vision-language-action model,

    Y . Wang, P. Ding, L. Li,et al., “Vla-adapter: An effective paradigm for tiny-scale vision-language-action model,” inProc. AAAI Conf. Artif. Intell., vol. 40, no. 22, 2026, pp. 18638–18646

    2026

  8. [8]

    Not all features are created equal: A mechanistic study of vision-language-action models,

    B. Grant, X. Zhao, P. Wang, “Not all features are created equal: A mechanistic study of vision-language-action models,”arXiv preprint arXiv:2603.19233, 2026

    arXiv 2026

  9. [9]

    When vision overrides language: Evaluating and mitigating counterfactual failures in VLA models,

    Y . Fang, Y . Feng, D. Jing,et al., “When vision overrides language: Evaluating and mitigating counterfactual failures in VLA models,” arXiv preprint arXiv:2602.17659, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    When would vision-proprioception policies fail in robotic manipulation?,

    J. Lu, W. Xia, Y . Wu,et al., “When would vision-proprioception policies fail in robotic manipulation?,” inProc. Int. Conf. Learn. Represent., 2026

    2026

  11. [11]

    Causality: Models, Reasoning, and Inference,

    J. Pearl, “Causality: Models, Reasoning, and Inference,”Cambridge University Press, 2009

    2009

  12. [12]

    Elements of causal inference: foundations and learning algorithms,

    J. Peters, D. Janzing, B. Sch ¨olkopf, “Elements of causal inference: foundations and learning algorithms,”MIT Press, 2017

    2017

  13. [13]

    Restoring linguistic grounding in VLA models via train-free attention recalibration,

    N. Zhang, B. Zhu, S. Zhou,et al., “Restoring linguistic grounding in VLA models via train-free attention recalibration,”arXiv preprint arXiv:2603.06001, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Towards deploying vla without fine- tuning: Plug-and-play inference-time vla policy steering via embodied evolutionary diffusion,

    Z. Li, J. Liu, Z. Dong,et al., “Towards deploying vla without fine- tuning: Plug-and-play inference-time vla policy steering via embodied evolutionary diffusion,”IEEE Robot. Autom. Lett., vol. 11, no. 5, pp. 6234–6241, 2026

    2026

  15. [15]

    Tag: Target-agnostic guidance for stable object-centric inference in vision-language-action models,

    J. Zhou, Z. Zhan, R. Zhai,et al., “Tag: Target-agnostic guidance for stable object-centric inference in vision-language-action models,” arXiv preprint arXiv:2603.24584, 2026

    arXiv 2026

  16. [16]

    Investigating gender bias in language models using causal mediation analysis,

    J. Vig, S. Gehrmann, Y . Belinkov,et al., “Investigating gender bias in language models using causal mediation analysis,” inAdv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 12388–12401

    2020

  17. [17]

    Causal abstraction: A theoretical foundation for mechanistic interpretability,

    A. Geiger, D. Ibeling, A. Zur,et al., “Causal abstraction: A theoretical foundation for mechanistic interpretability,”J. Mach. Learn. Res., vol. 26, no. 83, pp. 1–64, 2025

    2025

  18. [18]

    Counterfactual vqa: A cause- effect look at language bias,

    Y . Niu, K. Tang, H. Zhang,et al., “Counterfactual vqa: A cause- effect look at language bias,” inProc. IEEE Conf. Comput. Vis. Pattern Recog., 2021, pp. 12700–12710

    2021

  19. [19]

    Unbiased scene graph generation from biased training,

    K. Tang, Y . Niu, J. Huang,et al., “Unbiased scene graph generation from biased training,” inProc. IEEE Conf. Comput. Vis. Pattern Recog., 2020, pp. 3716–3725

    2020

  20. [20]

    Meta-causal learning for single domain generalization,

    J. Chen, Z. Gao, X. Wu,et al., “Meta-causal learning for single domain generalization,” inProc. IEEE Conf. Comput. Vis. Pattern Recog., 2023

    2023

  21. [21]

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

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

    Pith/arXiv arXiv 2025

  22. [22]

    Libero: Benchmarking knowledge transfer for lifelong robot learning,

    B. Liu, Y . Zhu, C. Gao,et al., “Libero: Benchmarking knowledge transfer for lifelong robot learning,” inAdv. Neural Inf. Process. Syst., vol. 36, 2023, pp. 44776–44791

    2023

  23. [23]

    Evaluating real-world robot manipulation policies in simulation,

    X. Li, K. Hsu, J. Gu,et al., “Evaluating real-world robot manipulation policies in simulation,”arXiv preprint arXiv:2405.05941, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Calvin: A benchmark for language-conditioned policy learning for long-horizon robot ma- nipulation tasks,

    O. Mees, L. Hermann, E. Rosete-Beas,et al., “Calvin: A benchmark for language-conditioned policy learning for long-horizon robot ma- nipulation tasks,”IEEE Robot. Autom. Lett., vol. 7, no. 3, pp. 7327– 7334, 2022

    2022

  25. [25]

    Univla: Learning to act anywhere with task-centric latent actions,

    Q. Bu, Y . Yang, J. Cai,et al., “Univla: Learning to act anywhere with task-centric latent actions,”arXiv preprint arXiv:2505.06111, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Unified vision-language-action model,

    Y . Wang, X. Li, W. Wang,et al., “Unified vision-language-action model,” inProc. Int. Conf. Learn. Represent., 2026

    2026

  27. [27]

    Discrete diffusion vla: Bringing dis- crete diffusion to action decoding in vision-language-action policies,

    Z. Liang, Y . Li, T. Yang,et al., “Discrete diffusion vla: Bringing dis- crete diffusion to action decoding in vision-language-action policies,” arXiv preprint arXiv:2508.20072, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Pd-vla: Accelerating Vision- Language-Action Model Integrated with Action Chunking via Parallel Decoding,

    W. Song, J. Chen, P. Ding,et al., “Pd-vla: Accelerating Vision- Language-Action Model Integrated with Action Chunking via Parallel Decoding,”arXiv preprint arXiv:2503.02310, 2025

    arXiv 2025

  29. [29]

    Molmoact: Action reasoning models that can reason in space,

    J. Lee, J. Duan, H. Fang,et al., “Molmoact: Action reasoning models that can reason in space,” inProc. IEEE Int. Conf. Robot. Autom., 2026

    2026

  30. [30]

    Thinkact: Vision-language-action reasoning via reinforced visual latent planning,

    C. Huang, Y . Wu, M. Chen,et al., “Thinkact: Vision-language-action reasoning via reinforced visual latent planning,” inAdv. Neural Inf. Process. Syst., vol. 38, 2026, pp. 82782–82802

    2026

  31. [31]

    Cot-vla: Visual chain-of-thought reasoning for vision-language-action models,

    Q. Zhao, Y . Lu, M. J. Kim,et al., “Cot-vla: Visual chain-of-thought reasoning for vision-language-action models,” inProc. IEEE Conf. Comput. Vis. Pattern Recog., 2025, pp. 1702–1713

    2025

  32. [32]

    Worldvla: Towards autoregressive action world model,

    J. Cen, C. Yu, H. Yuan,et al., “Worldvla: Towards autoregressive action world model,”arXiv preprint arXiv:2506.21539, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Smolvla: A vision- language-action model for affordable and efficient robotics,

    M. Shukor, D. Aubakirova, F. Capuano,et al., “Smolvla: A vision- language-action model for affordable and efficient robotics,”arXiv preprint arXiv:2506.01844, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Gr00t n1: An open foundation model for generalist humanoid robots,

    J. Bjorck, F. Castaneda, N. Cherniadev,et al., “Gr00t n1: An open foundation model for generalist humanoid robots,”arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Graspvla: A grasping foundation model pre-trained on billion-scale synthetic action data,

    S. Deng, M. Yan, S. Wei,et al., “Graspvla: A grasping foundation model pre-trained on billion-scale synthetic action data,” inProc. Conf. Robot Learn., 2025, pp. 1004–1029

    2025

  36. [36]

    Vla-os: Structuring and dissecting plan- ning representations and paradigms in vision-language-action models,

    C. Gao, Z. Liu, Z. Chi,et al., “Vla-os: Structuring and dissecting plan- ning representations and paradigms in vision-language-action models,” inAdv. Neural Inf. Process. Syst., vol. 38, 2026, pp. 136705–136736

    2026

  37. [37]

    Universal actions for enhanced embodied foundation models,

    J. Zheng, J. Li, D. Liu,et al., “Universal actions for enhanced embodied foundation models,” inProc. IEEE Conf. Comput. Vis. Pattern Recog., 2025, pp. 22508–22519

    2025

  38. [38]

    Qwen3-vl technical report,

    S. Bai, Y . Cai, R. Chen,et al., “Qwen3-vl technical report,”arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025