Architecture-Agnostic Test-Time Adaptation via Backprop-Free Embedding Alignment

This paper introduces a groundbreaking approach to test-time adaptation that operates without requiring backpropagation and works across different neural network architectures.

Abstract

Test-time adaptation has become crucial for deploying machine learning models in dynamic environments. Our work presents an architecture-agnostic method that achieves effective adaptation through embedding alignment techniques.

Key Contributions

• Architecture-agnostic adaptation framework
• Backpropagation-free optimization approach
• Superior performance across diverse model types
• Efficient computational requirements

Recommended citation: Xiao Ma, Young D. Kwon, Pan Zhou, Dong Ma. (2026). “Architecture-Agnostic Test-Time Adaptation via Backprop-Free Embedding Alignment.” ICLR 2026.