acceptodds
Withdrawn from ICLR 2027

VeriVLA: Adaptive Test-Time Verification for Flow-Based VLAs with Latent World Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Flow-based Vision-Language-Action (VLA) policies are inherently stochastic: different noise seeds can produce action chunks of varying quality from the same observation. At test time, however, the policy typically commits to a single sample without an explicit estimate of its quality. In fine-grained manipulation, executing a suboptimal action can irreversibly disrupt task progress and lead to failure. To exploit policy stochasticity, we introduce a training-free two-axis trigger that uses two initial samples to measure within-chunk temporal variation and cross-sample variance, detecting the critical states. Once triggered, a JEPA-based latent world model ranks candidate action chunks by predicting action-conditioned future latent states and the corresponding changes in task progress. By operating in latent space, the verifier avoids pixel reconstruction and enables efficient and more accurate ranking. Across three real-world tasks ( pen insertion, pipetting, and apple placement), our method raises the average completion score from 50.8% to 65.5%, while adding only 326 ms per triggered decision on a single RTX 4090. Consistent gains are also observed in simulation, where the average success rate rises from 66.4% to 73.6% in the clean setting of RoboTwin 2.0, from 62.8% to 69.6% in its randomized setting.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA, test-time scaling, JEPA

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aCBaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on test-time sampling for flow-based VLAs while adaptively deciding when additional sampling is needed. The authors propose VeriVLA, which uses a two-axis trigger based on temporal variation within action chunks and variance across two sampled chunks to decide when to use additional computation. When triggered, the proposed method samples more candidate actions and uses a JEPA-based latent world model to predict future latent states and delta task progress. It then executes the candidate with the highest predicted cumulative progress. The verifier is trained with Robometer-derived progress labels, while the VLA remains frozen at test time. The authors report an increase in average completion score from 50.8% to 65.5% on three real-world tasks, along with success-rate gains on RoboTwin 2.0 and LIBERO benchmarks.

Critical strengths
  • Clear problem setup and well-motivated approach: Flow-based VLAs are inherently stochastic when sampling actions, and sampling a suboptimal action may lead to task failure. The authors observe that failures tend to occur in high-variance segments and propose a trigger to identify these critical moments and allocate additional test-time sampling.

  • Trigger component is easy to understand and straightforward: The authors use variability in the policy’s sampled actions to identify potentially important moments by measuring temporal variation within each of two sampled action chunks and disparity between the chunks. These two axes provide a simple but quite reasonable signal for deciding when to perform additional sampling.

Critical weaknesses
  • The necessity of the trigger component is somewhat unclear: One of my main concerns is that I am not fully convinced why adaptive test-time verification is necessary, as the trade-off between performance gains and latency from additional test-time computation remains vague. Although the authors provide a trigger-rate sweep in Figure 6 and compare latency with always-on verification in Figure 5, they do not show the relationship among trigger rate, success rate, and latency within the same experimental setup.

  • The reliability of the delta-progress labels used to train the verifier is vague: I believe that test-time verification performance depends heavily on verifier quality. This paper uses absolute progress labels automatically generated by Robometer to train the latent world model to predict delta progress. However, the accuracy of these Robometer labels is not independently validated on the evaluated tasks, and it remains unclear whether they provide reliable supervision in various robot tasks and environments.

Questions for authors

Q1. Can the authors provide a performance-latency plot while varying the trigger threshold under the same experimental setup as the main Tables, including always-on verification? Figures 5 and 6 report latency and trigger-rate sweeps separately; a joint plot would help clarify the benefit of adaptive verification.

Q2. Can the authors examine the sensitivity of the verifier to the Robometer-derived progress labels? First, how does performance compare with alternative labels, such as other labeling methods or task-specific heuristics based on gripper states? Second, can the authors compare the Robometer progress labels with human-annotated progress labels on a validation dataset not used for training the verifier? Finally, do these labels remain reliable for higher-DoF embodiments, such as robot arms with dexterous hands or humanoids?

Q3. Can the authors provide an ablation of the progress loss comparing the Huber loss alone, the rank loss alone, and their combination? I would like to know whether the rank loss indeed contributes beyond the Huber loss.

Reviewer_uFfZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes test-time action selection for flow-based VLAs. A training-free trigger computes within-chunk temporal variation and cross-sample variance from two initial samples, then calibrates them as percentiles against base-policy rollouts. When the trigger fires, the policy draws K=10 candidates and a JEPA-based latent world model ranks them. The world model is initialized from V-JEPA 2-AC and adds a task-query progress head trained on Robometer labels. It predicts future latents and delta progress without decoding pixels. Results cover three real-world tasks on π0.5 (50.8 → 65.5), RoboTwin 2.0 on Hy-VLA, and LIBERO on π0.5 and StarVLA-GR00T.

Critical strengths
  1. The trigger is simple, cheap (14 ms), and practical. Its alignment with contact-critical states is evaluated against privileged simulator labels rather than only intuition.
  2. Verifying in latent space is a sensible way to avoid the latency of pixel world models, and the latency breakdown is clear.
  3. The real-world gains are meaningful, and the comparison with a pixel-WM verifier (GPC-Rank) and a VLM verifier (RoboMonkey) under the same delta-progress target is useful.
Critical weaknesses
  1. The verifier is trained on ordinary rollout transitions, not same-state counterfactual candidates (App. B), but at test time it must rank candidates from one state. The paper offers no direct evidence that the transfer works, such as same-state ranking accuracy against real outcomes.
  2. Progress labels come from Robometer, so the verifier's quality is bounded by Robometer's. It would be interesting to see how would it perform with GT progress labels.
  3. Related work misses WEAVER (Jain et al., 2026), which also uses a latent world model with reward and critic heads to select actions on π0.5 at test time. The novelty of a latent-WM verifier is weaker than claimed, and the adaptive trigger may be the main differentiator.
  4. The simulation experiments compare only against the base policy. All results use one seed, one model, and 50 trials. The LIBERO gain (96.9 → 97.6) and several RoboTwin gains are within a few episodes, and LIBERO is too saturated to say much.
  5. Missing ablations: k for the trigger (two samples give a noisy variance estimate), K for the verifier, and a random-selection baseline at matched K.
  6. In Table 4, removing world modeling leaves Huber loss nearly unchanged (0.26 vs. 0.25) but drops completion by 10 points. This needs explanation.
Questions for authors
  1. In simulation, can you report an oracle upper bound? This would mean restoring state, executing all K candidates, and selecting by ground-truth progress. Comparing it with VeriVLA, and with a variant trained on GT labels, would show whether the remaining gap comes from candidate diversity or from verifier error.
  2. How does the method compare with WEAVER and DreamAvoid?
  3. Can you evaluate on LIBERO-PRO or LIBERO-Plus, where the base policy fails more often?
  4. Can you report confidence intervals or multiple seeds for all benchmarks? And run the other baselines on the same benchmarks.
Reviewer_oPsGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The proposed VeriVLA compare action chunks from different noise seeds, use "within-chunk temporal variation" and "cross-sample variance" for better test time action verification to improve success rate. Specifically, with the two cues, the method uses JEPA-based latent world model as an external verifier to rank the action chunks and improve the success rate. The empirical validations report the proposed VeriVLA improves the success rate by noticeable margins with non-negligible latency (0.3 sec for each trigger).

Critical strengths
  • Practically useful method: Verification is only done when it is needed. So, it is suitable for efficient test-time scaling.
  • No pixel reconstruction required: Using JEPA-based world model, the verification is done in latent-space, which is much less expensive and practical.
  • Empirical improvement with not very significant latency (but the latency is not negligible)
Critical weaknesses
  • Novelty is limited: Action verification using latent world model has been done before (FOREWARN, Wu et al., 2025). Combining this idea with the idea of RoboMonkey (Kwok et al., 2025) that verifies candidates action chunks renders the proposed method. There is no new insight for solving this problem.
  • How to set the threshold? Should it be tuned for each task? If so, it is not truly "training free."
  • Not sufficient justification
    • Are two samples sufficient to estimate cross-sample variance? Do we need to have more than 2 to statistically meaningful estimation?
    • Within-chunk temporal variation may be caused by uncertainty or intentional rapid movement. Why is this signal corresponding to critical state of the simulator.
  • By the task definition of completion score of 0.5, it is hard to see if the improvement is coming from easy grasp or detailed maneuver for which the proposed method aims.
  • Presentation needs improvement
    • Fig 7's notation has no explanation, e.g., soft-crit and meaning of each frame screenshot. What is σ\sigma of Gauss? Is Gaussian distribution anyway used?
Questions for authors
  • How to set the threshold? Should it be tuned for each task?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.