SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models

Hyeonbeom Choi; Daechul Ahn; Youhan Lee; Taewook Kang; Seongwon Cho; Jonghyun Choi

Source-linked AI summary

SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models

Hyeonbeom Choi, Daechul Ahn, Youhan Lee, Taewook Kang, Seongwon Cho, Jonghyun Choi

arXiv:2602.04208v2Feb 4, 2026cs.ROcs.AIcs.LG

TL;DR

Existing VLA test-time scaling methods require extra training, verifiers, or repeated inference and generally modify action decoding without revisiting visual representations. SCALE uses self-uncertainty to jointly adapt visual attention and action sampling in one forward pass, consistently improving state-of-the-art VLAs and outperforming existing TTS methods. Its scope includes an assumption about correlated consecutive visual frames and broader deployment safety considerations.

Problem
Existing VLA test-time scaling methods require additional training or verifiers, multiple forward passes, and fixed visual representations that are insufficient under perceptual ambiguity.
Method
SCALE uses output-logit self-uncertainty to jointly modulate action sampling and vision-encoder attention without additional training, an external verifier, or multiple forward passes.
Results
SCALE consistently improves state-of-the-art VLAs across simulated and real-world benchmarks and outperforms existing TTS methods.
Takeaways & Limitations
SCALE broadens perception and action exploration under ambiguity while focusing execution when confident, enabling adaptive closed-loop control across varying conditions.
Takeaways & Limitations
Visual-attention adaptation assumes consecutive visual frames are highly correlated, using prior-step uncertainty as a proxy for the current timestep.

Abstract

from arXiv · show

Vision-Language-Action (VLA) models have emerged as a promising paradigm for general-purpose robotic control, with test-time scaling (TTS) gaining attention to enhance robustness beyond training. However, existing TTS methods for VLAs require additional training, verifiers, and multiple forward passes, making them impractical for deployment. Moreover, they intervene only at action decoding while keeping visual representations fixed-insufficient under perceptual ambiguity, where reconsidering how to perceive is as important as deciding what to do. To address these limitations, we propose SCALE, a simple inference strategy that jointly modulates visual perception and action based on 'self-uncertainty', inspired by uncertainty-driven exploration in Active Inference theory-requiring no additional training, no verifier, and only a single forward pass. SCALE broadens exploration in both perception and action under high uncertainty, while focusing on exploitation when confident-enabling adaptive execution across varying conditions. Experiments on simulated and real-world benchmarks demonstrate that SCALE improves state-of-the-art VLAs and outperforms existing TTS methods while maintaining single-pass efficiency.

1. Introduction

VLA inference can fail when fixed visual processing misses task-relevant cues or greedy decoding overlooks plausible actions. SCALE addresses these limitations by using self-uncertainty to jointly adapt perception and action, and it improves diverse VLA systems while retaining single-pass efficiency.

Autoregressive VLAs encode visual observations and sequentially decode action tokens conditioned on language instructions.
Existing TTS methods for VLAs require additional verifier training, can degrade under domain shift, and incur multiple forward passes.
Under perceptual ambiguity, action-only adaptation is insufficient because existing TTS methods keep visual representations fixed.
SCALE jointly modulates visual perception and action using self-uncertainty without additional training, an external verifier, or multiple forward passes.
SCALE combines a self-uncertainty measure with action sampling and visual-attention temperature adjustments to balance exploration and exploitation.
SCALE improves state-of-the-art VLAs across simulated and real-world benchmarks while maintaining single-pass efficiency.

2. Related Work

Related work has used uncertainty for decoding and test-time scaling, while visual-attention methods often rely on trained components. SCALE instead provides training-free, dynamic visual-attention modulation alongside adaptive action decoding.

Test-time scaling on VLA models: VLA test-time scaling commonly uses generate-and-verify strategies, including offline value functions, trained action verifiers, or self-verification.
SCALE overview: SCALE modulates the vision encoder’s attention temperature according to uncertainty deviation, sharpening focus when confident and broadening exploration when uncertain.
Uncertainty estimation in generative models: Prior uncertainty methods adapt decoding temperature, truncate distributions, or select reasoning paths, but VLA approaches have broader limitations addressed by SCALE.
Visual attention for VLMs and VLAs: Visual-attention research links task-relevant image-region allocation to VLM and VLA performance, but existing methods rely on contrastive masking or trained modules.
Visual attention for VLMs and VLAs: SCALE dynamically modulates visual attention during execution, broadening exploration under uncertainty and sharpening focus under confidence without additional training.

3. Approach

SCALE uses the VLA’s own predictive uncertainty to adapt both visual attention and action decoding, exploring under ambiguity and exploiting under confidence in a single pass.

Overview: SCALE jointly modulates visual perception and action from the model’s predictive uncertainty without additional training, external verifiers, or multiple rollouts.The procedure computes uncertainty from action-decoding logits and reuses the previous step’s uncertainty for visual modulation.
Preliminaries: Autoregressive VLA policies encode observations into visual representations and decode each action as a sequence of K tokens conditioned on language and prior tokens.The visual representation is produced by a Transformer-based vision encoder with attention temperature γ.
Motivation: Fixed greedy decoding and visual processing can struggle with perceptual ambiguity and action multimodality, where distractors or multiple plausible actions require broader exploration.SCALE therefore aims to broaden exploration under ambiguity while focusing execution when confidence is high.
Self-Uncertainty: SCALE positions the predicted token distribution between one-hot full certainty and uniform full ambiguity to derive a bounded self-uncertainty score from logits.The one-hot reference represents commitment to the top-1 token, while the uniform reference represents complete distributional uncertainty.
Adaptive Action Decoding: Adaptive action decoding maps token-level uncertainty through a sigmoid gate to sampling temperature, producing near-greedy execution under low uncertainty and exploratory sampling under high uncertainty.The temperature is scaled by T0, the maximum temperature defining the exploration range.
Adaptive Visual Attention: Adaptive visual attention averages token uncertainties, tracks their exponential moving average, and uses temporal deviation to adjust the vision encoder’s attention temperature.Vision-encoder modulation is chosen because it directly determines which visual information is extracted; the authors report it is more effective than modulating cross-modal attention.
Execution: The complete method maintains single-pass control by using previous-step uncertainty for visual modulation, avoiding additional rollouts, external verifiers, and auxiliary training.The assumption is that consecutive visual frames are highly correlated, and an additional-pass variant yields similar performance empirically.

4. Experiments

SCALE is evaluated across simulated and real-world benchmarks, multiple VLA backbones, and ablations against greedy, sampling, and training-required TTS baselines. It consistently improves performance, with especially large gains in real-world ID/OOD settings and when adaptive perception and decoding are combined.

Experimental setup: Experiments use OpenVLA, π0-FAST, and SpatialVLA across LIBERO, SIMPLER-WidowX, LIBERO-PRO-Long, and real-world ID/OOD tasks.Real-world evaluations use a 6-DoF UR10e arm and fine-tuned OpenVLA and π0-FAST models.
Quantitative results: SCALE improves over greedy decoding across all evaluated benchmarks and backbones, including +19.5/+13.9 ID gains and +16.7/+12.5 OOD gains for OpenVLA/π0-FAST.Simulation gains include +5.8 with OpenVLA on LIBERO and +14.6 with π0-FAST on SIMPLER-WidowX.
Quantitative results: Fixed temperature, top-k, and top-p sampling underperform greedy decoding on LIBERO with π0-FAST, whereas SCALE adapts sampling temperature to predicted uncertainty.The reported LIBERO scores change from 91.2 to 84.3, 88.1, and 86.2 for the three naive strategies.
Quantitative results: SCALE outperforms training-required TTS methods while remaining training-free and single-pass, exceeding MG-Select by +10.7 points on LIBERO with OpenVLA.It also surpasses RoboMonkey and TACO on LIBERO-Long by +6.8 and +3.3 points, respectively.
Ablations: Combining adaptive decoding and adaptive visual attention reaches 63.3, exceeding their summed individual gains by over 2%.Adaptive decoding contributes +5.3 and adaptive visual attention +3.3, while the combined result is +10.6 over the 52.7 baseline.
Ablations: The dual-reference uncertainty measure achieves the highest success rate among compared uncertainty measures, outperforming the next best by 5.5.Self-Certainty, which captures only distributional uncertainty, yields marginal gains.
Qualitative analysis: Adaptive visual attention broadens attention across the scene when uncertainty rises, redirecting focus from task-irrelevant regions toward the target mug.With fixed γt=1, the baseline attends to regions such as the microwave door and fails the task.

5. Conclusion

SCALE improves VLA robustness by jointly adapting perception and action according to self-uncertainty without additional training, an external verifier, or multiple forward passes. Its uncertainty measure captures both distributional concentration and top-1 confidence, and experiments show consistent improvements over SoTA VLAs and existing TTS methods.

Conclusion: SCALE jointly modulates perception and action from self-uncertainty using no additional training, no external verifier, and one forward pass.It broadens exploration under ambiguity and focuses on exploitation when confident.
Conclusion: The self-uncertainty measure compares both certainty extremes to capture distributional concentration and top-1 confidence.Experiments report consistent improvements over SoTA VLAs and existing TTS methods.

Impact Statement

The paper frames SCALE as a method for adaptive perception and action under uncertainty, with potential relevance to robot robustness, safety, and reliability. It also acknowledges broader deployment concerns and a self-referential uncertainty assumption.

Impact Statement: SCALE aims to improve embodied AI robustness by adapting perception and action under uncertainty.The stated motivation is to support robots operating across diverse real-world environments.
Impact Statement: The paper notes potential benefits and concerns from more capable robots, including improved automation, workforce displacement, and human-robot safety.It recommends safety protocols and human oversight for deployment.
Impact Statement: The low-uncertainty reference is self-referential because it anchors to the model’s own top-1 prediction rather than external ground truth.The metric is therefore designed to measure internal decisiveness, not externally verified correctness.
Impact Statement: The method’s uncertainty measure quantifies model conviction rather than prediction correctness.High conviction suggests less exploration, while diffuse predictions indicate ambiguity warranting broader exploration.
Impact Statement: Low average pmax is associated with significantly lower task success across 6,000 LIBERO episodes.This supports pmax as a signal of conviction and potential failure risk.
Impact Statement: SCALE uses preceding-step uncertainty to modulate visual attention while preserving single-pass inference.A two-step oracle achieves 64.6% but doubles evaluation time, whereas temporal correlation makes the preceding signal a proxy for current uncertainty.

E. Comparative Analysis of Inference and Training Efficiency

SCALE is evaluated against the inference latency and training overhead of test-time scaling methods. Its single-pass design avoids the increasing latency of multiple candidate generations and the additional training required by verifier-based approaches.

Evaluation Cost: The efficiency comparison uses wall-clock inference measurements and total evaluation time as cost indicators.Table 9 defines total evaluation time over 500 LIBERO-Long episodes on one NVIDIA A6000 GPU.
Inference Latency: Inference latency increases substantially as the number of generated action samples grows for OpenVLA and π0-FAST.At N = 16, latency increases by approximately 15.9× for OpenVLA and 3.2× for π0-FAST.
Training Overhead: Many TTS methods require auxiliary training for external verifiers, reward models, or self-verification.These requirements involve additional data collection and substantial training compute, limiting scalability and immediate deployment to unseen domains.
Inference Latency: SCALE maintains single-inference-pass efficiency instead of generating multiple action candidates.This avoids the latency costs associated with multi-sample test-time scaling.

F.1. Simulation Benchmarks

The evaluation spans simulation and real-world settings designed to test generalization, precise manipulation, robustness to perturbations, and seen or unseen objects. Experiments use multiple VLA backbones and model-specific implementation choices.

Simulation Benchmarks: The simulation evaluation covers LIBERO, SIMPLER-WidowX, and the challenging LIBERO-PRO-Long unseen split.These benchmarks target distribution-shift generalization, precise pick-and-place execution, and robustness beyond memorized trajectories.
Real-World Setup: The real-world setup uses a UR10e arm with third-person and wrist-mounted cameras.The experiments include three in-distribution and two out-of-distribution pick-and-place tasks, with OOD tasks introducing unseen object compliance or geometry.
VLA Backbones: The evaluation includes OpenVLA, π0-FAST, and SpatialVLA with distinct vision processing and action-tokenization schemes.OpenVLA predicts seven action tokens, π0-FAST uses FAST-compressed variable-length sequences, and SpatialVLA generates spatial tokens over a horizon of T = 4.
Implementation: SCALE is backbone-agnostic but adapts its implementation to each model’s vision encoder and action-tokenization scheme.Sampling and attention modulation are applied according to each backbone’s token structure, including factorized vocabularies and autoregressive propagation.
Evaluation Protocol: The experiments distinguish fine-tuned evaluation on seen tasks from zero-shot evaluation on LIBERO-PRO-Long.OpenVLA and π0-FAST use task-specific fine-tuning for standard benchmarks, while robustness evaluation uses LIBERO-Long-derived or zero-shot settings.

H. Sensitivity Analysis of Baseline Decoding Strategies

The sensitivity analysis finds that fixed-parameter decoding provides only marginal gains and varies little across hyperparameter settings. SCALE performs substantially better by adapting to changing predictive uncertainty.

Baseline Sensitivity: 76.2% to 77.2% is the narrow average-success range produced by varying temperature, top-k, and top-p settings.The fine-tuned OpenVLA baseline is 75.7%.
SCALE Comparison: 81.5% is SCALE’s average success rate, exceeding all fixed-parameter decoding strategies.The result supports the limitation of manually fixed hyperparameters across varying tasks and environmental states.
Interpretation: Fixed decoding settings cannot consistently adapt to different levels of predictive uncertainty.The analysis therefore selects representative temperature, top-k, and top-p configurations for the main comparisons.

I. Comparison with Test-Time Scaling Methods: Per-Task Breakdown

SCALE is compared with training-free and training-required test-time scaling methods on LIBERO-Long, alongside uncertainty-metric controls and fair implementation details. The comparison centers on per-task success rates using an OpenVLA backbone.

Per-Task Breakdown: Table 11 compares SCALE with greedy decoding and training-required test-time scaling methods on LIBERO-Long using an OpenVLA backbone.The benchmark contains long-horizon manipulation tasks requiring sustained precision over extended episodes.
Uncertainty Metrics: The uncertainty-metric comparison evaluates normalized entropy, confidence, Gini impurity, and Self-certainty alongside SCALE.All metrics are mapped to u ∈[0, 1], with 0 denoting maximum certainty and 1 maximum uncertainty.
Controlled Comparison: The baseline metrics replace σ(u_k) directly in SCALE’s action-decoding and visual-attention modulation while all other algorithmic components remain identical.This isolates the effect of the uncertainty metric on performance.

K. Statistical Reliability of Main Results

Across OpenVLA and π0-FAST, SCALE improves over fixed decoding baselines, with reported gains that remain larger than corresponding standard deviations. The results are reported over three random seeds and include a larger gain on LIBERO-Long.

OpenVLA: 81.5 ± 0.7% average success rate for OpenVLA exceeds Top-p sampling’s 77.2 ± 0.8%.On LIBERO-Long, SCALE reaches 63.3 ± 1.0% versus 55.2 ± 0.7% for the strongest baseline.
π0-FAST: 93.0±0.3% average success rate for π0-FAST exceeds its fine-tuned baseline’s 91.2 ± 0.4% across LIBERO suites.Fixed sampling-based decoding strategies often degrade performance relative to the fine-tuned baseline for π0-FAST.
Statistical Reliability: The main LIBERO results report means and standard deviations over three random seeds, with SCALE’s gains remaining larger than the corresponding standard deviations.This supports stability across random seeds rather than reliance on a single favorable run.

L. Hyperparameter Sensitivity Analysis

SCALE is generally robust to tested choices of T0, α, and ϵ, while κ is more sensitive because it directly controls visual-attention temperature. Hyperparameters selected on LIBERO-Long are reused across other benchmarks and real-world experiments.

T0 and α: All tested T0 and α values outperform greedy baselines for both OpenVLA and π0-FAST, indicating robustness across these settings.The corresponding greedy baselines are 52.7% for OpenVLA and 76.3% for π0-FAST.
Base Temperature: OpenVLA performs best at T0=1.0, whereas π0-FAST performs best at T0=0.3, consistent with differences in action vocabulary size and tokenization.π0-FAST uses |V|=2048, while OpenVLA uses |V|=256.
Visual Attention Bound: κ is more sensitive than T0 and α: OpenVLA drops from 63.3% at κ=2.0 to 39.8% at κ=3.0, while π0-FAST drops from 80.9% to 74.4%.κ controls γ_t ∈(1/κ, κ), and excessive values can over-flatten visual attention.
Reference Smoothing: 57.0%–58.0% success rates across ϵ values from 10^-10 to 10^-14 all exceed the greedy baseline by at least 4.3 percentage points.This sensitivity test uses OpenVLA with only adaptive action decoding on LIBERO-Long.
Cross-Benchmark Transfer: Using one LIBERO-Long-selected configuration across simulation and real-world experiments, SCALE consistently improves performance without benchmark-specific retuning.The same configuration is applied to all other simulation benchmarks and real-world experiments.