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

The gist: Flow-based Vision-Language-Action (VLA) policies are inherently stochastic, and this paper introduces VeriVLA, an adaptive test-time verification framework that uses a training-free two-axis trigger and a JEPA-based latent world model to rank candidate action chunks by predicting action-conditioned future latent states and the corresponding changes in task progress.

How it works

The method addresses the issue where different noise seeds produce samples of varying quality during inference, which can irreversibly disrupt task progress in finegrained manipulation The paper introduces 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.

Two-Axis Trigger

The trigger is designed to detect critical moments by combining two complementary signals (C1) activate in contact-rich regions, where manipulation is inherently more complex and failures concentrate, and (C2) activate at states where repeated sampling produces diverse candidate actions, providing the verifier with meaningful alternatives to choose from. The temporal axis measures the mean squared changes between consecutive actions within each sampled chunk, defined as Ttemp(st) = 1/k X k i=1 1 H − 1 H X−1 j=1 a˜(i)t+j − a˜(i)t+j−12 The cross-sample axis measures cross-sample action variance: Tcross(st) = 1/k X k i=1 1 H H X−1 j=0 a˜(i)t+j − a¯˜t+j 2 The overall trigger score is then defined as τ (st) = maxτtemp(st), τcross(st) and it follows an OR rule, activating when either percentile exceeds the threshold: τ (st) > τ0.

Latent World Model Verifier

When the Two-Axis Trigger is activated, the verifier Vϕ evaluates the quality of action chunk samples by predicting delta progress without the need for a goal image The verifier uses a JointEmbedding Predictive Architecture (JEPA) world model framework, which consists of a pretrained visual encoder and an action-conditioned autoregressive predictor. The architecture introduces embodiment-specific state encoders and action encoders together with sub-chunk rollouts to provide denser supervision while reducing inference overhead. A delta progress prediction branch is introduced, which estimates how much the executed action sub-chunk advances the specified task, requiring a task-specific learnable query token to condition delta progress prediction The verifier is jointly optimized for latent prediction and progress estimation, with progress estimation as the primary objective.

Results and Evaluation

Experiments across three real-world tasks (pen insertion, pipetting, and apple placement) show consistent gains On three real-world tasks, our method improves the average completion score from 50.8% to 65.5% while adding 326 ms for sampling and verification per triggered decision on a single RTX 4090 In simulation, 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, and from 96.9% to 97.6% on LIBERO. Ablation studies confirm the benefits of latent dynamics modeling and pretraining, showing that without the dense world modeling objective, the model struggles to learn informative dynamics

Limitations and Future Work

A key limitation of repeated sampling and selection is the difficulty of steering a pretrained VLA model Another limitation is the use of a frozen visual encoder, as these representations are not explicitly optimized for manipulation progress estimation Future work can investigate learning an encoder–predictor pair for progress estimation within a JEPA framework by considering progress estimation as an optimization goal.

Conclusion

We have introduced an adaptive test-time verification framework for exploiting the stochasticity of flow-based VLA policies Our training-free two-axis trigger identifies critical states from within-chunk motion variation and cross-sample disagreement, invoking additional sampling only when it is likely to be useful A JEPA based latent world model then ranks candidate action chunks by jointly predicting action-conditioned future latents and delta task progress, avoiding pixel reconstruction and goal-image dependence Experiments in both real-world and simulation demonstrate consistent gains across tasks, environments, and base policies. On three real-world tasks, our method increases the average completion score from 50.8% to 65.5% while adding 326 ms for sampling and verification per triggered decision, excluding the 14 ms trigger overhead Ablation studies further confirm the benefits of latent dynamics modeling and pretraining Taken together, these results establish adaptive latent verification as an effective and practical approach to improving frozen generative VLAs at deployment.

1. Bold Header: Adaptive Test-Time Verification for Flow-Based VLAs

Exploit policy stochasticity by introducing 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. This allows the system to bypass expensive pixel reconstruction and instead use a JEPAbased latent world model to rank candidate action chunks by predicting action-conditioned future latent states and the corresponding changes in task progress.

2. Bold Header: Real-Time Latency Reduction via Latent Verification

The system can execute actions with a total latency of 422 ms when triggered, but verification only adds 326 ms for sampling and verification per triggered decision on a single RTX 4090, which is significantly lower than pixel-level world models that incur substantial latency, making it difficult to deploy at real-time control frequencies.

3. Bold Header: Improved Fine-Grained Manipulation Performance

The system can raise the average completion score from 50.8% to 65.5% across three real-world tasks (pen insertion, pipetting, and apple placement), demonstrating a 14.7 percentage points improvement over the base policy through adaptive verification at critical decision steps.

4. Bold Header: Task-Specific Progress Estimation

The system can perform task-specific progress estimation by introducing a task-specific learnable query token to condition delta progress prediction, allowing the model to distinguish between different task semantics, such as where the same state transition corresponds to positive progress for the task 'closing the closet', but negative progress for 'opening the closet'.

5. Bold Header: Robustness in Randomized Environments

The verifier can boost base policy success rates in randomized settings by 6.8 percentage points on RoboTwin 2.0 Clean2Random, showcasing a generalization capability where the model can effectively select high-quality actions from diverse stochastic samples.

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, i.e., pen insertion, pipetting, and apple placement), our method raises the average completion score from 50.8% to 65.5%, while adding 326 ms for sampling and verification 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, and from 96.9% to 97.6% on LIBERO.