Research map / Vision-language and multimodal / Robot action and navigation

Enabling VLA Action Self-Verification via VLM Token Probability Bucketing

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TokenPB, a self-verification method that co-trains a VLA's VLM backbone with a bucket-token likelihood loss alongside a flow-matching action expert, then selects the best of M sampled action chunks by teacher-forced bucket-token scoring.

Problem

Test-time sampling improves VLA policies only if a good candidate action can be selected efficiently; existing selectors rely on separately trained external verifiers or per-candidate scoring passes that add overhead.

Approach

Discretize continuous action chunks into bucket tokens and train the VLM backbone with an auxiliary cross-entropy bucket loss alongside flow matching, so the same checkpoint scores sampled candidates via teacher-forced log-likelihood with shared observation-prefix KV-cache reuse.

Key results

• TokenPB improves Qwen2.5-VL-GR00T LIBERO average from 93.9 to 96.8 and pi0.5 from 96.5 to 98.4 at M=16 over single-sample inference.
• On real-world tasks TokenPB gains +17.1 points on average (seen) and +16.4 on unseen settings relative to single-sample inference.
• TokenPB keeps extra inference time below 0.2 s at M=64, versus over 4 s for RoboMonkey and 1.5 s for MG-Select at M=64.

Surprising findings

• Random selection among candidates gives no gain over single-sample inference, so candidate diversity alone is useless without a scoring function.
• Adding the bucket loss during training alone already improves LIBERO success rate by about 2 points, suggesting co-training does not hurt flow-matching.

Limitations

• GPU memory still grows linearly with the number of candidates M, which may limit edge deployment.
• The verifier favors frequent demonstration actions over globally optimal ones, and its guarantees depend on an unverified alignment between demonstration likelihood and task utility.
• Action types absent from training demonstrations cannot be scored meaningfully.

Experimental setup

Datasets: LIBERO CALVIN RoboTwin 2.0 SimplerEnv-WidowX

Benchmarks: LIBERO CALVIN ABC->D RoboTwin 2.0 SimplerEnv-WidowX real-world bimanual tasks

Metrics: success rate average length action RMSE inference time

Baselines: RoboMonkey MG-Select AsyncVLA RoVer V-GPS MoDE random selection majority voting

Builds on: GR00T-N1 flow matching FAST action tokenizer RoboMonkey MG-Select

Closest papers by meaning

• VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.90)
• VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.89)
• VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.89)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning (CVPR 2026 · similarity 0.88)