Preprint claims small-parameter refinement cuts audio deepfake detection error
CoReLoop reportedly cuts pooled cross-domain EER from 4.85% to 3.74% with ~10M trainable parameters, but the result is self-tested and awaits independent replication.
Original event 2026-09-17
A preprint named CoReLoop reports that 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.
The method leaves the original model's parameters and training data untouched, reusing encoder outputs through lightweight refinement modules and low-rank adapters; an optional halting head picks refinement depth per utterance, reaching 3.73% pooled EER at an average of 1.18 passes.
The results are first-party, self-run benchmarks; the composition of the 14 test sets and the identity of the baseline detector need checking in the full text, and real-world generalization awaits independent replication.