Research map / Vision-language and multimodal / Robot action and navigation
Trust Your Instincts: Confidence-Driven Test-Time RL for Vision-Language-Action Models
Chen Siyao, Jiakang Yuan, Jiaxin Wang, Tao Chen
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 |
|---|---|---|---|
| apply T2VLA self-reward test-time RL | OpenVLA-OFT · LIBERO average | success rate: +6.2 | Improved |
| apply T2VLA self-reward test-time RL | pi0 · LIBERO average | success rate: +24.2 | Improved |
| apply T2VLA self-reward test-time RL | pi0.5 · LIBERO average | success rate: +8.0 | Improved |
| apply T2VLA self-reward test-time RL | OpenVLA-OFT · RoboTwin 2.0 average | success rate: +21.3 | Improved |
| add the global expert pool to local expert only (ablation) | T2VLA · LIBERO-Long | success rate: 94.5 to 95.3 | Improved |
| add the local pseudo-expert to global pool only (ablation) | T2VLA · LIBERO-Long | success rate: 93.0 to 95.3 | Improved |
| set pool capacity K=5 instead of 3 (ablation) | T2VLA · LIBERO-Long | success rate: 93.4 to 95.3 | Improved |
| set pool capacity K=5 instead of 10 (ablation) | T2VLA · LIBERO-Long | success rate: 91.4 to 95.3 | Improved |
| use min-max confidence normalization instead of static weighting (ablation) | T2VLA · LIBERO-Long | success rate: 88.3 to 95.3 | Improved |
| use min-max confidence normalization instead of max routing (ablation) | T2VLA · LIBERO-Long | success rate: 87.5 to 95.3 | Improved |
| use min-max normalization instead of sigmoid gate (ablation) | T2VLA · LIBERO-Long | success rate: 92.6 to 95.3 | Improved |
| apply T2VLA to a 1-shot SFT policy | OpenVLA-OFT · LIBERO-Long (1-shot) | success rate: 17.3 to 11.0 | Hurt |
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning (CVPR 2026 · similarity 0.90)
- Enabling VLA Action Self-Verification via VLM Token Probability Bucketing (NeurIPS 2026 · similarity 0.90)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.90)
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.
