Research map / Vision-language and multimodal / Robot action and navigation
Verifier-free Test-Time Sampling for Vision Language Action Models
Suhyeok Jang, Dongyoung Kim, Changyeon Kim, Youngsuk Kim, Jinwoo Shin
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 MG-Select test-time selection to the base model without joint training (ablation) | pi0-FAST · RoboCasa pick-and-place (100 demos) | success rate (%): 17.0 -> 22.6 | Improved |
| add joint training (condition dropout) without MG-Select (ablation) | pi0-FAST · RoboCasa pick-and-place (100 demos) | success rate (%): 17.0 -> 28.5 | Improved |
| add MG-Select on top of joint training (ablation) | pi0-FAST with joint training · RoboCasa pick-and-place (100 demos) | success rate (%): 28.5 -> 31.0 | Improved |
| select by KL to condition-masking reference instead of by likelihood (ablation) | MG-Select (N=4) · RoboCasa pick-and-place (100 demos) | success rate (%): 30.5 -> 31.0 | Improved |
| increase number of candidates N from 4 to 64 (ablation) | MG-Select · RoboCasa pick-and-place (100 demos) | success rate (%): 31.0 -> 33.3 | Improved |
| mask text only versus text and state together (ablation) | MG-Select · RoboCasa pick-and-place (100 demos) | success rate (%): 29.7 -> 31.0 | Improved |
| raise regularization temperature of masking reference from 1.0 to 4.0 (ablation) | MG-Select · RoboCasa pick-and-place (100 demos) | success rate (%): 28.8 -> 31.0 | Improved |
| truncate aggregation to first 5 tokens instead of summing all tokens (ablation) | MG-Select · RoboCasa pick-and-place (100 demos) | success rate (%): 26.1 -> 31.0 | Improved |
| joint-training dropout ratio 10/10/10 instead of 5/5/5 (ablation) | MG-Select with joint training · RoboCasa pick-and-place (100 demos) | success rate (%): 24.2 -> 31.0 | Improved |
| add single-prefill deployment sharing one prefill across candidates | MG-Select · LIBERO-Object latency (N=4) | latency (s): 43.4 -> 23.7 | Improved |
| truncate to first 5 tokens instead of summing all tokens (ablation) | MG-Select* · SIMPLER-WidowX (average of 4 tasks) | success rate (%): 49.0 -> 50.3 | Mixed |
| replace RoboMonkey external verifier with MG-Select confidence | pi0-FAST with N=4 · SIMPLER-WidowX (average of 4 tasks) | success rate (%): 50.7 -> 50.3 | No clear effect |
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- Enabling VLA Action Self-Verification via VLM Token Probability Bucketing (NeurIPS 2026 · similarity 0.91)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.91)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.91)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 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.
