Enables model adaptation on edge devices and non-differentiable (quantized) models using a purely backpropagation-free optimization framework.
March 17, 2026
Original Paper
ZOTTA: Test-Time Adaptation with Gradient-Free Zeroth-Order Optimization
arXiv · 2603.14254
The Takeaway
Existing test-time adaptation (TTA) requires high-memory backpropagation, which is often impossible on edge hardware or with hard-quantized weights. ZOTTA uses Zeroth-Order Optimization with automated layer selection to achieve adaptation performance comparable to BP-based methods using only forward passes.
From the abstract
Test-time adaptation (TTA) aims to improve model robustness under distribution shifts by adapting to unlabeled test data, but most existing methods rely on backpropagation (BP), which is computationally costly and incompatible with non-differentiable models such as quantized models, limiting practical deployment on numerous edge devices. Recent BP-free approaches alleviate overhead but remain either architecture-specific or limited in optimization capacity to handle high-dimensional models. We p