acceptodds
Under review as a conference paper at ICLR 2027

Vision-Language-Action Models are Inherent Verifiers: Test-Time-Scaling via Likelihood-Based Visual Prompting

Abstract

Test-time scaling (TTS) has recently emerged as a powerful paradigm for improving Vision-Language-Action (VLA) models without requiring expensive fine-tuning. However, existing TTS approaches typically rely on training external verifiers or are strictly limited to autoregressive action decoders. To address these limitations, we introduce Likelihood-Based Visual Prompting (LVP), a training-free, verifier-free test-time scaling framework and reveal that generalist VLA policies can serve as their own inherent verifiers. LVP first projects action candidates directly into the 2D image space via perspective-aware action projection, overlaying future trajectory and gripper poses as structured visual prompts. We then reframe action verification as an evaluative Visual Question Answering (VQA) task. By prompting the VLA to evaluate the quality of the visually prompted action via a multiple-choice question (e.g., grading the action from "Poor" to "Excellent"), we leverage the model's internal conditional token likelihoods to compute an expected action quality score for candidate ranking. Furthermore, we introduce an attention-guided token pruning mechanism to alleviate the computational overhead of processing multi-sample visual prompts. By leveraging the visual attention maps generated during the action proposal stage, LVP dynamically discards task-irrelevant visual tokens during the verification phase. In the experiments, LVP improves autoregressive and diffusion/flow-matching based VLA models across diverse simulation and real-world benchmarks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Test-Time Scaling, Robotic Manipulation

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_wUmFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LVP (Likelihood-Based Visual Prompting), a training-free test-time scaling framework for VLA models. Candidate actions are projected onto the observation image as translation trajectories and rotation axes, then ranked by the expected utility of multiple-choice VQA answers computed from conditional token probabilities of the language-modeling head. The authors also propose a visual token pruning method that combines attention from the action generation stage with feature differences before and after visual overlay, reducing verification latency. The framework is evaluated on both autoregressive and flow-matching policies.

Critical strengths
  1. The method turns action candidates into visual prompts and scores them with the model's existing language-modeling capability, without training a dedicated verifier. This converts action evaluation into a visual-semantic judgment and supports both autoregressive and flow-matching action decoders within the same framework, which gives it broad applicability.

  2. The experiments are solid: evaluation covers multiple backbones, benchmarks, and real-robot tests, with a range of ablations.

Critical weaknesses
  1. The visual action representation omits gripper open/close. Sec. 3.2 and Appendix A.2 define candidate actions with translation, rotation, and gripper state, yet the visual prompt shows only the translation trajectory and goal orientation. Candidates with the same end-effector trajectory and orientation but different gripper commands thus receive identical visual representations, even though gripper timing directly affects grasping and releasing success. The paper should clarify which action dimensions the score covers and how gripper differences participate in candidate selection.

  2. No direct quantitative validation links candidate scores to execution quality. While Tab. 5 reports final success rates, it does not verify whether higher-scoring candidates succeed more often than lower-scoring ones within the same pool. This gap relates to the performance drop as the candidate pool grows: in Tab. 10a, GR00T's full-token average success rate on WidowX drops from 60.0 at N=4 to 56.0 at N=16, and the pruned version drops from 58.4 to 54.0, below the base policy's 55.0. Appendix A.7 attributes this drop to verification noise without identifying its cause. A candidate-level analysis would clarify in which states the score is reliable and why larger pools select actions with lower execution quality.

Questions for authors

Is the gripper open/close command included in the verifier input? When candidates differ in grasp or release timing, how do the visual prompt and scoring function distinguish them?

On a small set of representative simulation states, could you execute the candidate action chunks from the same pool with a fixed follow-up policy and compare LVP scores with downstream success? Showing the gap between LVP's selection and the oracle selection in each pool would help explain why performance drops at large N.

2 comments

We thank you for the constructive feedback and insightful suggestions. The reviewer highlights the strength of our training-free, decoder-agnostic paradigm, solid cross-backbone/benchmark evaluations, and real-robot validations. Below, we address each weakness and question in detail.

(Weakness 1 & Question 1: Gripper command representation) Is the gripper open/close command included in the verifier input? When candidates differ in grasp or release timing, how do the visual prompt and scoring function distinguish them?

Reply: Thank you for pointing out this important detail. We do not use the gripper open/close command as the verifier input, and mainly focus on the geometric alignment of action candidates. We defined action candidates with translation, rotation, and gripper states in Sec. 3.2 and Appendix A.2, because the VLA policies jointly generate these candidates. We clarify this detail in Appendix A.2 of the revised manuscript.

(Weakness 2 & Question 2: Candidate score vs. execution quality validation) On a small set of representative simulation states, could you execute the candidate action chunks from the same pool with a fixed follow-up policy and compare LVP scores with downstream success? Showing the gap between LVP's selection and the oracle selection in each pool would help explain why performance drops at large N.

Reply: We appreciate this valuable experimental suggestion. To directly validate whether higher candidate scores correlate with downstream execution success and to analyze the performance gap at large pool sizes (N=16N=16), we conduct an empirical candidate-level execution experiment.

Experimental Setup

  • Backbone & Benchmark: GR00T N1.7 on SimplerEnv (WidowX).
  • Evaluation Protocol: We select 10 representative simulation states. For each state, we sample a pool of N=16N=16 action candidates and execute individual candidates using a vanilla GR00T N1.7 backbone across 8 random seeds.
  • Baselines:
    1. LVP (Ours): Soft expected utility over 4 choices.
    2. LLaVA-7B Baseline: LVP using external VLM. LLaVA-7B is a base VLM of RoboMonkey's verifier.
    3. Binary VQA: P(Yes)P(Yes) scoring over binary choice. (D) variant of Tab. 5.
    4. Argmax Choice Ablation: Hard discrete token selection. (E) variant of Tab. 5.
  • Reference Metrics: Across all candidate pools, the mean candidate success rate is 56.6%, and the Oracle success rate (selecting a candidate with the best downstream success rate in each pool) is 86.3%.

Experimental Results

Verification MethodArgmax Selection Success RateMean Success (Score≥0.5\text{Score} \ge 0.5)Mean Success (Score<0.5\text{Score} < 0.5)Avg. Score Ties per Pool (N=16N=16)
Mean Candidate Pool56.6%---
Oracle selection86.3%---
LLaVA-7B baseline51.3%57.1%56.4%1.00
Binary VQA ablation53.8%56.5%54.3%3.00
Argmax choice ablation53.4%--14.82
LVP (Ours)55.0%57.6%56.1%1.18

Analysis

  1. Positive Score-to-Quality Correlation: Candidates assigned high normalized scores (Score≥0.5\text{Score} \ge 0.5) under LVP achieve a higher mean success rate (57.6%) compared to low-scoring candidates (56.1% for Score<0.5\text{Score} < 0.5) and the candidate pool average (56.6%). This confirms that LVP's likelihood score reliably correlates with ground-truth downstream execution quality.
  2. Tie-Breaking Analysis: While binary VQA and argmax choice ablations suffer from large score ties (3.00 and 14.82), LVP (Ours) reduces score ties to 1.18, proving that our continuous soft expected utility over localized token probabilities is essential for fine-grained trajectory ranking.
  3. Oracle Gap and Large NN Behavior: The gap between LVP's top selection (55.0%) and the Oracle selection (86.3%) illustrates that while likelihood verification provides a strong zero-shot ranking signal, soft likelihoods are not perfect ground-truth verifiers. As the candidate pool expands (N=16N=16), the chance of sampling noisy false positives increases. However, LVP's continuous scoring preserves significantly higher selection quality than alternative verifiers. We expect that replacing strict argmax selection with soft sampling over the verification scores will reduce the probability of selecting noisy false positives, making test-time scaling more stable as NN grows.
AuthorsOpenReview ↗

Gentle follow-up regarding our rebuttal

Dear Reviewer wUmF,

Thank you again for your time and constructive feedback. Please let us know if our responses and additional experiments have addressed your concerns, or if you have any further questions. We look forward to any further feedback you might have.

Best regards, The Authors

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

The paper tests the hypothesis of whether the VLM part of a VLA can be used to verify its own actions using visual prompting (draw VLA trajectories onto a 2D image and let the VLM grade each, then the best is selected). Their claim is that VLAs possess "inherent self-verification" capabilities that can be used to verify itself through the aforementioned strategy. They also provide a token pruning step that makes the verification faster and evaluate their method on two autoregressive and two flow matching VLAs. While individual results are sometimes worse they show gains on average across all settings

Critical strengths

The approach and assumed cause of the findings is interesting beyond the question in the paper. Can VLAs self-verify their own trajectories? I believe the implication could lead to interesting follow up research on self-verification given that it turns out to be true

Critical weaknesses

To actually verify whether the measured gains come from the VLA grading its own candidates, a comparison with different VLMs grading the same VLA trajectories in images is necessary. If the gains turn out to be equivalent, the claim of "inherent self-verification" does not hold.

Furthermore, the claim would collapse into "VLMs can be used to select among candidate actions drawn into the image," which PIVOT (Nasiriany et al., 2024) has already shown. I would expect such an experiment in the paper.

Questions for authors

The main claim is the "inherent self-verification" and I think this does not hold yet from the experiments provided. What would change my rating is if you could show that different VLMs consistently perform worse. Softening the claim would not be enough to increase the score as I think the work would not be novel enough.

2 comments

We thank you for the insightful review and for highlighting the core significance of our work: "Can VLAs self-verify their own trajectories? I believe the implication could lead to interesting follow-up research." Below, we present new empirical results directly addressing this question, alongside a detailed clarification of our novelty relative to PIVOT (Nasiriany et al., 2024).

1. Inherent Self-Verification vs. External VLM Verifiers

To validate whether the performance gains stem from inherent self-verification rather than generic VLM visual grading, we conduct comparative experiments across OpenVLA and MolmoAct. We compare success rate of three strategies:

  1. Base policy: single-pass baseline execution.
  2. LVP (Ours, Self-Verification): Trajectory selection using the VLA policy's native pre-trained vision-language backbone.
  3. External VLM Baseline: LVP without token pruning using an off-the-shelf VLM instead of the VLA policy. For fair comparison, we select the base VLM models (Prismatic-7B for OpenVLA, Molmo-7B for MolmoAct).

Experimental Results

Table R1. OpenVLA Backbone on LIBERO-Long Benchmark

Method / VerificationToken PruningSuccess RateGain over Base
OpenVLA-52.7%-
External VLM VerifierFull tokens58.8%+6.1%
LVP (Ours)Full tokens61.2%+8.5%
LVP (Ours)Pruned, M′=16M'=1661.0%+8.3%

Table R2. MolmoAct Backbone on SimplerEnv (WidowX) Benchmark

TaskBase PolicyExternal VLM VerifierLVP (Full tokens)LVP (Pruned, M′=28M'=28)
Pick coke can (Zero-Shot / Fine-tuned)71.3% / 77.7%81.7% / 89.0%82.7% / 87.0%82.7% / 91.3%
Move Near (Zero-Shot / Fine-tuned)73.8% / 77.1%72.2% / 80.6%75.9% / 77.8%70.4% / 75.9%
Open/Close drawer (Zero-Shot / Fine-tuned)66.5% / 60.0%73.1% / 73.1%78.7% / 75.0%76.4% / 77.3%
Average (Zero-Shot / Fine-tuned)70.5% / 71.6%75.7% / 80.9%79.1% / 80.4%76.0% / 81.5%

As shown in Tab. R.1 and R.2, LVP achieves superior success rates compared to external VLM verifiers across evaluated benchmarks. While VLM models exhibit general-purpose visual reasoning, VLA backbones acquire robotics-aligned perception through domain-specific action supervision, developing spatial alignment to manipulation affordances. Consequently, inherent VLA self-verification provides a more accurate verification signal for its own candidate proposals while reusing the model's existing parameters without external models and enabling efficient visual token pruning.

2. Conceptual and Algorithmic Distinctions from PIVOT (Nasiriany et al., 2024)

We further clarify three fundamental differences between LVP and PIVOT:

  • Candidate Sampling Source: PIVOT samples candidate actions from unconditioned random or heuristic Gaussian distributions, leading to low candidate efficiency. LVP samples candidates directly from the generalist VLA policy's learned action distribution πθ(a∣I,g)\pi_{\theta}(a|I,g), which is a reasonable prior action distribution.
  • Unified Single-Model Architecture: PIVOT requires proprietary VLM evaluators (e.g., GPT-4V), introducing heavy computational load or API cost. LVP unifies proposal generation and likelihood evaluation within a single VLA model, which is much more efficient than these VLM models.
  • Verification Signal: PIVOT projects multiple action candidates into image and queries VLM to pick the best candidate. On the other hand, LVP extracts soft expected utility over localized next-token log-probabilities (S(ai)=∑vkW(vk)⋅Pθ(vk∣Ii,q)\mathcal{S}(a_i) = \sum_{v_k} \mathcal{W}(v_k) \cdot P_\theta(v_k \mid I_i, q)), enabling continuous, well-calibrated trajectory ranking.
AuthorsOpenReview ↗

Gentle follow-up regarding our rebuttal

Dear Reviewer iYuA,

Thank you again for your time and constructive feedback. Please let us know if our responses and additional experiments have addressed your concerns, or if you have any further questions. We look forward to any further feedback you might have.

Best regards, The Authors

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

This paper introduces Likelihood-Based Visual Prompting (LVP), a training-free test-time scaling method for VLA policies. LVP projects sampled actions onto the observation image, uses the frozen VLA's conditional label likelihoods to rank them through evaluative VQA, and applies attention-guided token pruning to reduce verification cost. Experiments across autoregressive and flow-matching models in simulation and real-world tasks show consistent success-rate gains.

Critical strengths

1.LVP reframes action verification as visual-language evaluation, enabling training-free candidate selection for both autoregressive and flow-matching VLA policies without an external verifier.

2.The method is evaluated across several VLA backbones, simulation benchmarks, and real-world tasks; token pruning also provides a 2.26x verification speedup at N = 16.

Critical weaknesses

1.The paper reports the best result for each setting but does not state whether N was selected on a separate validation set. Simulation trial counts, random seeds, and uncertainty estimates are also missing.

2.Some reported results are inconsistent. In Table 7, the pruned row combines the Spatial, Goal, and Long scores from the (N=8) setting in Table 9 with the Object score from (N=4), and its four suite scores average 78.85% rather than the reported 78.4%. The full-token result of 78.1% is also below the 79.6% reported for (N=16) in Table 9, despite the statement that the best result is reported for each setting.

3.For GR00T, verification loads the full Cosmos-Reason2-2B model although the policy uses only early VLM layers, leaving the shared weights and the scope of the self-verification claim unclear.

4.The hard-label ablation does not specify tie-breaking, and the binary-baseline result does not establish the claimed inability to distinguish subtle quality differences. RoboMonkey's degradation may also reflect candidate-distribution mismatch rather than overfitting.

5.The pruning budget is ambiguous because the union of two selected token sets may exceed the stated count. The reported latency is for OpenVLA rather than the GR00T real-world setup, whose end-to-end replanning latency is not reported.

Questions for authors

please refer to critical weaknesses

3 comments

We thank you for the constructive feedback and insightful suggestions. The reviewer highlights our novel reframing of action verification as visual-language evaluation, broad cross-backbone/benchmark evaluations, real-robot validations, and the effectiveness of our token pruning speedup. Below, we address each weakness and question in detail.

(Weakness 1: Hyperparameter selection (NN), trial counts, and uncertainty estimates) The paper reports the best result for each setting but does not state whether NN was selected on a separate validation set. Simulation trial counts, random seeds, and uncertainty estimates are also missing.

Reply:

  1. Selection of Candidate Pool Size (N): While we report the best result for each setting in Tab. 2, Tab. 3, and Tab. 7 (Tab. 8 in revised manuscript), and we provide the full results across N∈4,8,16N \in {4,8,16} in Tab. 9, Tab. 10, and Tab. 11 (Tab. 10, Tab. 11, and Tab. 12 in revised manuscript). Based on these full empirical results, N=4N=4 offers an optimal trade-off between task accuracy and inference efficiency, which we adopt as the standard setting for Tab. 1 and Tab. 4 without separate validation sets (We follow the convention to evaluate VLA policies under simulation benchmark or real-world setup [R1]). We clarify this detail in Appendix A.7, Tab. 10, Tab. 11, and Tab. 12 of the revised manuscript.

[R1] Kim et al., Openvla: An open-source vision-language-action model. In Conference on Robot Learning (CoRL), 2025.

  1. Simulation Trial Counts: Our evaluation involves extensive simulation rollouts across all benchmarks to ensure statistical reliability:
  • MolmoAct on SimplerEnv (WidowX) in Tab. 1: Evaluated over 624 total trials, comprising
    • pick_coke_can: 300 trials (3 subtasks ×\times 4 perturbations ×\times 25 rollouts)
    • move_near: 108 trials (4 perturbations ×\times 27 rollouts)
    • open/close drawer: 216 trials (6 subtasks ×\times 4 perturbations ×\times 9 rollouts)
  • GR00T N1.7 on SimplerEnv in Tab. 2: Evaluated over 200 trials per task across 7 tasks (Spoon, Carrot, Eggplant, Cube, Open drawer, Close drawer, Place in closed drawer), total 1,400 trials.
  • LIBERO-Plus-Long Benchmark in Tab. 3: Evaluated over 2,519 total trials.
  • LIBERO Benchmark in Tab. 7 (Tab. 8 in revised manuscript): Evaluated over 2,000 total trials (4 task suites ×\times 10 tasks/suite ×\times 50 rollouts per task).
  1. Random seeds and Uncertainty Estimates: We are currently running the experiments over multiple random seeds. We will update the results with standard deviations.

(Weakness 2: Inconsistencies of table) Some reported results are inconsistent across tables. In Table 7, the pruned row combines scores from different NN settings in Table 9, and the average math (78.85% vs 78.4%) is misaligned. Full-token results also show discrepancies.

Reply: Thank you for pointing out these numerical discrepancies. We fix 88.6% to 86.6% (Object, the pruned row) in Tab. 7 (Tab. 8 in revised manuscript), which matches the reported average 78.4%. In the main quantitative evaluation, we select NN that shows the best performance under both a full token budget and our pruned token budget. For example, N=4N=4 shows average 78.1%, N=8N=8 shows average 78.25% and N=16N=16 shows average 78.25% over the full token and pruned token budget in Tab. 9 (Tab. 10 in revised manuscript). We update Tab. 8 and clarify how to select NN for the main tables in Appendix A.7 of the revised manuscript.

(Weakness 3: GR00T architecture and self verification scope) For GR00T, verification loads the full Cosmos-Reason2-2B model although the policy uses only early VLM layers, leaving the shared weights and the scope of the self-verification claim unclear.

Reply: Our likelihood-based verification in Sec. 3.3 requires computing next-token likelihoods via a language modeling head. We can share the early VLM layers on GR00T N1.7 and Cosmos-Reason2-2B if the shared weights are the same. However, the officially released checkpoints have different parameters, so we load the full model for verification. Even though we load the original VLM for GR00T backbone, we do not load any external models for the other experiments (OpenVLA, MolmoAct, and MolmoAct2).

(Weakness 4: Clarification of ablations) The hard-label ablation does not specify tie-breaking, and the binary-baseline result does not establish the claimed inability to distinguish subtle quality differences. RoboMonkey's degradation may reflect candidate-distribution mismatch.

Reply:

  1. Hard-Label Argmax Tie-Breaking and Candidate-Level Analysis: We conducted a detailed candidate-level execution experiment on GR00T N1.7 (Simpler WidowX) across 10 representative states (N=16N=16 candidate pool, 8 random seeds per candidate, 1,280 rollouts). We kindly refer the reviewer xgny to the response of the reviewer wUmF for the details and results of this experiment.

    • Tie-Breaking Rule: When multiple candidates receive identical scores, we choose the earlier generated action candidate.
    • Distinguish Subtle Quantity Differences: While binary VQA and argmax choice ablations suffer from large score ties (3.00 and 14.82), LVP (Ours) reduces score ties to 1.18, proving that our continuous soft expected utility over localized token probabilities is essential for fine-grained trajectory ranking.
  2. RoboMonkey Degradation Analysis: The performance degradation of RoboMonkey in Tab. 2a comes from a candidate-distribution mismatch between training and test time. As mentioned in Appendix A.6, the action verifier of RoboMonkey is trained on action candidates generated by OpenVLA. When evaluated at test time on candidate pools generated by GR00T N1.7, the verifier fails to generalize to the candidate-distribution mismatch, resulting in miscalibrated scores and degraded candidate selection. This degradation highlights a key motivation of LVP: avoiding reliance on specialized verifiers trained on policy-specific action distributions. As shown across our experiments, LVP consistently yields performance gains across diverse policy architectures (OpenVLA, MolmoAct, MolmoAct2, GR00T N1.7). We clarify this candidate distribution shift analysis in the revised manuscript.

(Weakness 5: Pruning Budget and Real-World Latency) The pruning budget is ambiguous because the union of two selected token sets may exceed the stated count. End-to-end replanning latency for GR00T in the real-world setup is missing.

Reply:

  1. Pruning Budget: Taking the union of two selected token sets yields an actual evaluated visual token count Munion=∣Xattn∪Xprompt∣M_{\text{union}} = |\mathbf{X}^{\text{attn}} \cup \mathbf{X}^{\text{prompt}}| bounded by: M′≤Munion≤2M′M' \le M_{\text{union}} \le 2M' Because the robot arm and target object frequently overlap with the rendered visual prompt trajectory, a substantial portion of salient tokens is shared across both sets. The actual evaluated token count averages Munion=40.6±1.4M_{\text{union}} = 40.6 \pm 1.4 tokens using MolmoAct backbone on the SimplerEnv (Google Robot), where M=392M=392 and M′=28M'=28.

  2. End-to-End Real-World Latency: We have added end-to-end latency for real-world setup (N=4N=4):

  • Vanilla: 66.047 ± 7.248 ms
  • LVP (full token budget, M=64M=64): 140.136 ± 10.401 ms
  • LVP (pruned token budget, M′=8M'=8): 127.697 ± 15.163 ms (~ 8.0 Hz)

Since GR00T predicts action chunks executed over 15 control steps, and end-to-end latency of 127.697 ms fits comfortably within closed-loop execution horizons, enabling real-time deployment.

AuthorsOpenReview ↗

Gentle follow-up regarding our rebuttal

Dear Reviewer xgny,

Thank you again for your time and constructive feedback. Please let us know if our responses and additional experiments have addressed your concerns, or if you have any further questions. We look forward to any further feedback you might have.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.