Sensitivity of backpropagation-free adaptation to compression

Investigate whether backpropagation-free test-time adaptation methods exhibit sensitivity to structured compression that differs from the sensitivity of backpropagation-based methods.

Background

The main diagnostics rely primarily on gradients and therefore focus on backpropagation-based test-time adaptation. The paper also evaluates backpropagation-free methods such as FOA and PEA through accuracy curves, but does not extend the representational and gradient-based diagnostic framework to them. Consequently, whether backpropagation-free adaptation mechanisms respond to structured compression through different failure modes or sensitivities remains unresolved.

References

First, although our empirical evaluation covers both backpropagation-based and backpropagation-free TTA approaches (FOA~\citep{niu2024test}, PEA~\citep{xiao2026architectureagnostic}), our diagnostics are mainly gradient-based and do not yet cover backpropagation-free adaptation methods; we leave this as an open direction. The corresponding accuracy curves are reported in Fig.~\ref{fig:bp_free_hero}: the benefit of backpropagation-free adaptation likewise collapses toward the no-adaptation baseline as compression increases.

— On the Interaction Between Model Compression and Test-Time Adaptation  (2609.03604 - Corti et al., 3 Sep 2026) in Section 6, “Discussion, Outlook and Limitations,” subsection “Outlook”