Research map / Vision-language and multimodal / Robot action and navigation
Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment
Jacky Kwok, Xilun Zhang, Mengdi Xu, Yuejiang Liu, Azalia Mirhoseini, Chelsea Finn, Marco Pavone
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CoVer verifier test-time selection | pi0 · SIMPLER in-distribution (red-team instructions) | success rate avg (%): 41.5 -> 57.0 | Improved |
| add CoVer verifier test-time selection | pi0 · SIMPLER out-of-distribution (red-team instructions) | success rate avg (%): 29.7 -> 61.0 | Improved |
| add CoVer on top of rephrase-augmented policy | pi0 (rephrase) · SIMPLER in-distribution (red-team instructions) | success rate avg (%): 44.0 -> 65.5 | Improved |
| use CoVer vs training-time instruction augmentation (same data) | pi0 · SIMPLER out-of-distribution (red-team instructions) | success rate avg (%): 48.7 -> 61.0 | Improved |
| verify with CoVer vs random rephrase without verification (ablation) | pi0 · SIMPLER out-of-distribution (red-team instructions) | success rate avg (%): 28.7 -> 61.0 | Improved |
| use CoVer instead of RoboMonkey verifier | pi0 · SIMPLER in-distribution (red-team instructions) | success rate avg (%): 7.5 -> 57.0 | Improved |
| add CoVer to pi0 on Zucchini on Towel (OOD) | pi0 · SIMPLER Zucchini on Towel (OOD) | success rate (%): 30 -> 41 | Improved |
| add CoVer to pi0 on Carrot on Plate | pi0 · SIMPLER Carrot on Plate (in-distribution) | success rate (%): 48 -> 48 | No clear effect |
| scale sampled action candidates from 1 to 16 with CoVer (ablation) | CoVer-VLA · Bridge V2 held-out action error | action RMSE: 0.166 -> 0.147 | Improved |
| use instruction rephrasing instead of repeated sampling or Gaussian perturbation (ablation) | action proposal · Bridge V2 test-time scaling | oracle action error | Improved |
| scale rephrases and actions jointly (hybrid sampling) (ablation) | action proposal · Bridge V2 test-time scaling | oracle action error | Improved |
| use verification instead of scaling policy training (training compute) | CoVer · training cost | relative FLOPs: 16.0x -> 3.8x | Improved |
Closest papers by meaning
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.91)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.91)
- Enabling VLA Action Self-Verification via VLM Token Probability Bucketing (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
