Research map / Vision-language and multimodal / CLIP and zero-shot
Test-Time Perturbation Tuning with Delayed Feedback for Vision-Language-Action Models
Zehua Zang, Xi Wang, Fuchun Sun, Xiao Xu, Lixiang Liu, Jiahuan Zhou, Jiangmeng Li
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 delayed feedback perturbation learning on top of voting (ablation) | PDF (OpenVLA) · LIBERO-Object | success rate: 0.50 -> 0.72 | Improved |
| add data augmentation voting to perturbation learning (ablation) | PDF (OpenVLA) · LIBERO-Object | success rate: 0.67 -> 0.72 | Improved |
| add data augmentation voting to perturbation learning (ablation) | PDF (OpenVLA) · LIBERO-Long | success rate: 0.57 -> 0.59 | Improved |
| add delayed feedback to voting only (ablation) | PDF (OpenVLA) · LIBERO-Spatial | success rate: 0.80 -> 0.89 | Improved |
| voting without delayed feedback versus plain OpenVLA (ablation) | OpenVLA · LIBERO-Object | success rate: 0.62 -> 0.50 | Hurt |
| increase test-time augmentation budget (ablation) | PDF · LIBERO and Atari | performance | Hurt |
| add KL regularizer to REINFORCE term (ablation) | PDF · Atari-57 | human normalized score: 1.04 -> 1.07 | Improved |
| add REINFORCE term to KL term (ablation) | PDF · LIBERO-Object | success rate: 0.69 -> 0.72 | Improved |
| use dimension-wise voting instead of action-wise voting (ablation) | PDF · LIBERO | success rate: 0.68 vs 0.88 (Black Bowl); 0.86 vs 0.64 (Alphabet Soup) | Mixed |
| apply PDF test-time adaptation to Jat | Jat · Atari-57 | human normalized score: 0.97 -> 1.07 | Improved |
| apply PDF to OpenVLA | OpenVLA · LIBERO average | success rate: 0.69 -> 0.77 | Improved |
| apply PDF to OpenVLA on Object suite | OpenVLA (paper-reported) · LIBERO-Object | success rate: 0.86 -> 0.72 | Hurt |
Closest papers by meaning
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.89)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- TTRV: Test-Time Reinforcement Learning for Vision Language Models (CVPR 2026 · similarity 0.88)
- Test-Time Adaptation for Online Vision-Language Navigation with Feedback-based Reinforcement Learning (ICML 2025 · similarity 0.88)
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.
