10M-parameter refinement reportedly cuts cross-domain audio deepfake error

Training only about 10 million trainable parameters (out of 598 million) on an already-trained SSL-based audio deepfake detector cuts pooled equal error rate from 4.85% to 3.74% across 14 cross-domain test sets.

Previously, improving cross-domain detection typically meant altering the original model's parameters or adding training data; CoReLoop leaves both untouched, reusing encoder outputs through lightweight refinement modules and low-rank adapters.

The 4.85%-to-3.74% pooled EER reduction is first-party, self-run by the authors; an optional halting head picks refinement depth per utterance, reaching 3.73% pooled EER at an average of 1.18 passes.

The composition of the 14 test sets and the identity of the baseline detector need checking in the full text, real-world generalization awaits independent replication, and the preprint was submitted on 2026-09-17.