JEPA · 2026-10-16
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
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
VeriVLA uses a training-free two-axis trigger to decide when to spend extra compute on repeated sampling for flow-based VLA policies. When triggered, a JEPA-based latent world-model verifier ranks candidate action chunks by predicting action-conditioned future latent states and estimated delta progress, avoiding pixel reconstruction. The authors report improved completion/success scores on real-world tasks (pen insertion, pipetting, apple placement) and simulation benchmarks, at added verification latency.
Why it matters
Fine-grained manipulation can fail irreversibly if a stochastic VLA samples a poor action chunk. VeriVLA targets this by adaptively invoking verification only at critical states, then ranking candidates using latent dynamics and delta progress prediction. This aims to improve closed-loop task performance while reducing unnecessary inference cost compared to always-on verification and avoiding latency from pixel-based world models.
Method
- Two-Axis Trigger: training-free trigger using two sampled action chunks to measure temporal action variation and cross-sample variance; activates repeated sampling when either percentile exceeds a threshold.
- Latent World Model Verifier: a JEPA-based latent world model ranks candidate action chunks by predicting latent future and task-specific delta progress (via a task query token).
- Latency optimizations: KV-cache reuse, pipeline overlap, and PyTorch-level optimization to reduce added latency per triggered decision.
Abstract (from arXiv)
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, $\textit{i.e.,}$ 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.
Related papers
- PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models
- MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation
- FastOPD: On-Policy Distillation for Lightweight VLA Deployment
- Fast-VLA: You Only Think Once for Robot Manipulation
- EvoMem-VLA: State-Evolution Memory for Long-Horizon Robot Manipulation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-15 · Indexed by Pub-AI: 2026-10-16 UTC
VeriVLA: Adaptive Test-Time Verification for Flow-Based VLAs with Latent World Models
Preprint posted: 2026-10-15 · Original record ↗
- arxiv
- 2610.21985
Explicit source versions (1)
- v1 · 2026-10-15
- Preprint available
Source date: 2026-10-15 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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, $\textit{i.e.,}$ 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.