MAEB

Null space audio trick pushes embedding model to top of benchmark

A clever zero-training trick just pushed a 577M parameter embedding model to the top of the MAEB audio benchmark, outperforming models 15 times its size. By projecting intermediate Conformer layers into the unused null space of the final projection matrix, the macro score jumped 5.1 points to 54.4%. The approach preserves text embeddings bit-for-bit while unlocking untapped acoustic signal. It's a reminder that even overtrained foundation models hold hidden potential, and that sometimes the simplest interventions yield the most surprising results.