arxiv_cv 95% Match Research Paper Machine Learning Researchers,Computer Vision Engineers,AI Practitioners 2 weeks ago

SteeringTTA: Guiding Diffusion Trajectories for Robust Test-Time-Adaptation

computer-vision › diffusion-models

📄 Abstract

Abstract: Test-time adaptation (TTA) aims to correct performance degradation of deep models under distribution shifts by updating models or inputs using unlabeled test data. Input-only diffusion-based TTA methods improve robustness for classification to corruptions but rely on gradient guidance, limiting exploration and generalization across distortion types. We propose SteeringTTA, an inference-only framework that adapts Feynman-Kac steering to guide diffusion-based input adaptation for classification with rewards driven by pseudo-label. SteeringTTA maintains multiple particle trajectories, steered by a combination of cumulative top-K probabilities and an entropy schedule, to balance exploration and confidence. On ImageNet-C, SteeringTTA consistently outperforms the baseline without any model updates or source data.

Authors (5)

Jihyun Yu

Yoojin Oh

Wonho Bae

Mingyu Kim

Junhyug Noh

Submitted

October 16, 2025

arXiv Category

cs.CV

arXiv PDF

Key Contributions

SteeringTTA introduces an inference-only framework that adapts Feynman-Kac steering to guide diffusion-based input adaptation for classification using pseudo-labels. It maintains multiple particle trajectories steered by cumulative probabilities and an entropy schedule to balance exploration and confidence, leading to improved robustness against distribution shifts without model updates.

Business Value

Enhances the reliability of AI models in real-world scenarios where data distributions can change unexpectedly, reducing the need for costly retraining or manual intervention.

Paper Metadata

Innovation Type

Algorithmic

Deployment Feasibility

High, as it's an inference-only framework and doesn't require model retraining.

Limitations Addressed

Limitations of existing diffusion-based TTA methods that rely on gradient guidance, limiting exploration and generalization across distortion types. It also addresses the need for adaptation without model updates or source data.

Performance Gains

Consistently outperforms baseline on ImageNet-C.

Technical Tags

test-time adaptationdiffusion modelsinput adaptationFeynman-Kac steeringpseudo-labelingparticle trajectoriesentropy schedulerobustnessclassificationImageNet-C

Research Topics

Robustness in Deep LearningDistribution Shift AdaptationGenerative Models for AdaptationInput-based Model AdaptationUnsupervised Learning

Methods & Architectures

Feynman-Kac steeringdiffusion-based input adaptationpseudo-labelingparticle filtering Diffusion Models

Applications & Tasks

Computer Vision Image Classification Distribution ShiftModel Robustness DegradationGeneralization to Corruptions Image ClassificationTest-Time Adaptation

Datasets & Benchmarks

Datasets

ImageNet-C

Accuracy

Related Fields

Machine LearningDeep LearningComputer VisionGenerative Models

Keywords

Test-Time AdaptationDiffusion ModelsFeynman-KacSteeringInput AdaptationPseudo-labelingRobustnessDistribution ShiftImage ClassificationDeep LearningUnsupervised LearningGenerative AI

Academic Context

#Robustness in Deep Learning#Distribution Shift Adaptation#Generative Models for Adaptation#Input-based Model Adaptation#Unsupervised Learning

Commercial Potential

Potential Products

Robust image classification systemsAdaptive AI services

Target Industries

TechnologyHealthcareAutomotive

Use Case Examples

Adapting image classifiers to different camera sensors or lighting conditionsImproving robustness of medical image analysis under varying acquisition parameters

Competitive Edge

Offers an input-only adaptation method that leverages diffusion models and steering principles, aiming for better generalization than gradient-guided diffusion TTA.

Resource Requirements

Compute Needs

Inference-time compute for diffusion models.

Data Requirements

Requires a test dataset for adaptation.

Deployment Constraints

Performance might depend on the quality of pseudo-labels and the chosen entropy schedule.

Scalability

Inference-only nature suggests good scalability for adaptation.

Production Readiness

Maturity Level

Research

Time to Market

Medium

View Full Paper Back to Papers