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

Xueli Sun, Yuxiang Lu, Yantong Liu, Xianzhe Fan, Xi Chen, Zhe Liu^†, Hengshuang Zhao^*

The University of Hong Kong

^* Corresponding author^† Project lead

Flow-based VLAs are stochastic by design. We introduce a latent world model verifier to evaluate candidate action chunks by predicting their future states and task progress.

+14.7%

Real-world completion score

50.8% → 65.5%

+7.2%

RoboTwin 2.0 Clean success

66.4% → 73.6%

326 ms

Sampling + verification overhead

Per triggered decision RTX 4090

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 (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.

Figure 1: action variance concentrates in contact-rich regions in LIBERO, and the proposed trigger tracks critical task states. — Repeated sampling reveals higher action variance in contact-rich regions, where task failures are most likely to occur. Our proposed trigger value aligns with task difficulty and the policy's internal uncertainty. These observations motivate triggering repeated sampling and verification at critical decision steps.

Paradigm comparison

We argue that latent world models are a natural fit in the setting of repeated sampling and verification. The verifier evaluates the quality of action chunk samples by predicting the progress or relative ranking. Therefore, the model must understand world dynamics and representation of the future states resulting from executing the action chunk. In addition, verification is a low-dimensional problem: how much an action chunk advances the task is a scalar judgment, and most camera pixels are irrelevant to it. This motivates verification in a latent space, which compresses the pixel observation into compact and dynamic-rich features, where we can jointly predict the latent states and task progress.

Method

Two-axis trigger

Repeated sampling and verification add computation, so the trigger allocates this cost to states where it is most useful. It is designed to capture two complementary conditions: contact-rich interactions, where manipulation is more complex and failures tend to occur, and uncertain states, where different samples provide meaningful alternatives for the verifier.

Two initial action chunks provide both signals. The temporal axis measures changes between consecutive actions within each chunk, capturing motion variation associated with challenging interactions. The sample axis measures disagreement between the two chunks, capturing policy uncertainty. Each chunk is normalized by its average action magnitude to avoid favoring fast-moving stages.

Latent world model verifier

Action selection requires estimating the consequences of each candidate. A JEPA-based world model predicts action-conditioned future visual latents and task progress in a shared latent space, avoiding the cost of reconstructing future images. Once triggered, the policy expands the pool to 10 candidates, and the verifier selects the action chunk with the largest cumulative predicted progress.

Each action chunk is divided into sub-chunks for multi-step latent prediction. This provides denser supervision than predicting only the final state, while reducing inference cost and error accumulation compared with predicting after every individual action.

Progress estimation also requires task information: the same screen-lifting action can advance “opening the computer” while reversing “closing the computer.” A task-specific learnable query conditions the progress prediction branch, allowing the verifier to evaluate future states relative to the instructed task.

The attention mask prevents latent state prediction from attending to the task query. Without this restriction, task information could leak into the dynamics model and encourage shortcuts based on task semantics rather than the physical consequences of the action. The progress branch can use task information to judge whether a transition is useful, while the latent prediction branch models the environment dynamics independently of the task. This separation adapts the pretrained world model for verification while preserving its original modeling objective.

Task query and attention mask

Opening the computer gives positive progress for the screen-lifting action. The mask keeps task information out of latent dynamics prediction.

Real-world manipulation tasks

We evaluate the method on AgileX PiperX for three tasks, including pen insertion, pipetting, and apple placement.

PiperX robot grasping and inserting a pen into a narrow bottle.

Pen insertion

Grasp a pen and insert it into a yogurt bottle.

45.0%60.0%

Dual-arm robot transferring liquid between bottles using a pipette.

Pipetting

Transfer liquid from the right bottle to the left.

52.5%63.5%

Robot placing a grasped apple into a container.

Apple placement

Pick up an apple and place it in a container.

55.0%73.0%

Evaluation results

Inference efficiency

We implement KV caching, parallel execution and inference optimizations to reduce triggered sampling and verification overhead from 1,261 ms to 326 ms.

Citation

@misc{sun2026verivla,
  title  = {VeriVLA: Adaptive Test-Time Verification for
            Flow-Based VLAs with Latent World Models},
  author = {Sun, Xueli and Lu, Yuxiang and Liu, Yantong and
            Fan, Xianzhe and Chen, Xi and Liu, Zhe and
            Zhao, Hengshuang},
  year   = {2026},
  note   = {Preprint}
}