AG-REPA: Causal Layer Selection for Representation Alignment in Audio Flow Matching

Researchers introduce AG-REPA, a new method for improving audio generation models by strategically selecting which neural network layers to align with teacher models. The approach identifies that layers storing the most information aren't necessarily the most important for generation, leading to better performance in speech and audio synthesis.

• →AG-REPA introduces causal layer selection for better audio flow matching model training.
• →Research reveals Store-Contribute Dissociation where information-rich layers don't necessarily drive generation most effectively.
• →Forward-only gate ablation quantifies each layer's contribution to the velocity field that controls generation.
• →Method consistently outperforms baseline REPA across different audio training scenarios.
• →Results show alignment works best on causally dominant layers rather than just representationally rich ones.