Research map / Large language models / Reasoning and chain of thought

VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

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.

Adds an uncertainty-triggered test-time deliberation phase to flow-matching VLAs (PI0, PI0.5), using a tournament over sampled action chunks scored by a lightweight pairwise Relative Action Critic trained on automatically curated preference pairs from ODE-step quality differences.

Problem

Fast, reflexive VLA action generation fails in complex or ambiguous situations, while existing parallel deliberation methods apply costly critic-based selection indiscriminately and rely on unstable absolute action scoring.

Approach

Gate deliberation with a DTW-based disagreement score between two sampled action chunks (a "cognitive clutch"), share the VLM prefill across candidates, and select the best of N candidates via single-elimination pairwise comparisons by a Relative Action Critic that reuses VLM features through learnable query tokens and multi-branch attention; preference pairs are generated automatically by contras

Key results

• On LIBERO-LONG, VLA-ATTC raises PI0.5 average success from 90.6% to 95.4% (full) and PI0 from 82.8% to 92.2% (full), reducing PI0.5 failure rate by over 50%.
• On real-world Agilex Piper tasks, VLA-ATTC (full) improves PI0 average success from 46% to 63.3% (+17.3%).
• With the cognitive clutch, control frequency is 20.8 Hz versus 23.3 Hz for the baseline and 1.5 Hz for RoboMonkey.

Surprising findings

• Lowering the uncertainty threshold (triggering more deliberation) does not significantly improve success, suggesting few states are truly difficult.
• The uncertainty measure uses only two samples with DTW, which matched human agreement (89.2%) better than four or eight samples with mean pairwise distance (90.1%, 90.4%), so extra samples bring little gain.

Limitations

• Relies on a single uncertainty threshold set by percentile over offline data and gating only helps in sparse difficult states.
• Real-world evaluation is limited to three tasks on one robot arm.
• Full-deliberation variant reduces control frequency to 12.1 Hz, so per-step cost is substantial when always active.

Experimental setup

Datasets: LIBERO-LONG Agilex Piper real-world demonstrations

Benchmarks: LIBERO-LONG Agilex Piper real-world tasks

Metrics: task success rate control frequency (Hz) human agreement (%)

Baselines: PI0 PI0.5 RoboMonkey

Builds on: PI0 PI0.5 RoboMonkey flow matching Best-of-N sampling

Closest papers by meaning

• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
• VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
• VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.92)
• ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.92)
• NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
• ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.91)