LATTA: Langevin-Anchored Test-Time Adaptation for Enhanced Robustness and Stability

2025-10-06

AI Summary

Existing test-time adaptation (TTA) methods—such as Tent—exhibit instability and catastrophic forgetting under small batch sizes or severe data corruption. To address this, we propose NoiseAnchor: a lightweight, label-free, architecture-agnostic, and sampling-free TTA method. Its core innovation is stochastic weight perturbation inspired by stochastic gradient Langevin dynamics (SGLD), where source-model parameters serve as stable anchors to enforce Bayesian-style regularization, thereby mitigating overfitting to the loss landscape. Evaluated on Rotated-MNIST and CIFAR-10-C, NoiseAnchor consistently outperforms Tent, CoTTA, and EATA—achieving over a 2% improvement in average accuracy on CIFAR-10-C while significantly reducing performance variance. These results demonstrate superior robustness and stability under challenging test-time distribution shifts.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Reasoning under Uncertainty: Stochastic Optimization

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

Test-time adaptation (TTA) aims to adapt a pretrained model to distribution shifts using only unlabeled test data. While promising, existing methods like Tent suffer from instability and can catastrophically forget the source knowledge, especially with small batch sizes or challenging corruptions. We argue that this arises from overly deterministic updates on a complex loss surface. In this paper, we introduce Langevin-Anchored Test-Time Adaptation (LATTA), a novel approach that regularizes adaptation through two key mechanisms: (1) a noisy weight perturbation inspired by Stochastic Gradient Langevin Dynamics (SGLD) to explore the local parameter space and escape poor local minima, and (2) a stable weight anchor that prevents the model from diverging from its robust source pre-training. This combination allows LATTA to adapt effectively without sacrificing stability. Unlike prior Bayesian TTA methods, LATTA requires no architectural changes or expensive Monte Carlo passes. We conduct extensive experiments on standard benchmarks, including Rotated-MNIST and the more challenging CIFAR-10-C. Our results demonstrate that LATTA significantly outperforms existing methods, including Tent, CoTTA, and EATA, setting a new state of the art for self-supervised TTA by improving average accuracy on CIFAR-10-C by over 2% while simultaneously reducing performance variance.

Problem

• Enhancing robustness against distribution shifts in test-time adaptation
• Preventing catastrophic forgetting of source knowledge during adaptation
• Improving adaptation stability with noisy updates and weight anchoring

Innovation

• Uses noisy weight perturbation for local exploration
• Employs stable weight anchor to prevent divergence
• Requires no architectural changes or expensive passes