Guardado en:
Detalles Bibliográficos
Autores principales: Staniszewski, Łukasz, Zaleska, Katarzyna, Modrzejewski, Mateusz, Deja, Kamil
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:https://arxiv.org/abs/2602.11910
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866917508713283584
author Staniszewski, Łukasz
Zaleska, Katarzyna
Modrzejewski, Mateusz
Deja, Kamil
author_facet Staniszewski, Łukasz
Zaleska, Katarzyna
Modrzejewski, Mateusz
Deja, Kamil
contents Audio diffusion models can synthesize high-fidelity music from text, yet achieving fine-grained control over specific musical attributes remains challenging, as their internal mechanisms for representing high-level concepts are poorly understood. In this work, we use activation patching to demonstrate that recent audio diffusion architectures exhibit a semantic bottleneck, where a small, shared subset of consecutive attention layers controls distinct musical concepts, such as the presence of specific instruments, vocals, or genres. Building on this, we systematically evaluate a broad spectrum of steering paradigms, comparing activation steering against prompt-level, score-space, and weight-space interventions, analyzing the interaction between the steering mechanism and the intervention site. Our new benchmark, supported by an extensive user study, demonstrates that localized activation steering establishes a new state-of-the-art in audio concept modulation.
format Preprint
id arxiv_https___arxiv_org_abs_2602_11910
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle TADA! Tuning Audio Diffusion Models through Activation Steering
Staniszewski, Łukasz
Zaleska, Katarzyna
Modrzejewski, Mateusz
Deja, Kamil
Sound
Machine Learning
Audio diffusion models can synthesize high-fidelity music from text, yet achieving fine-grained control over specific musical attributes remains challenging, as their internal mechanisms for representing high-level concepts are poorly understood. In this work, we use activation patching to demonstrate that recent audio diffusion architectures exhibit a semantic bottleneck, where a small, shared subset of consecutive attention layers controls distinct musical concepts, such as the presence of specific instruments, vocals, or genres. Building on this, we systematically evaluate a broad spectrum of steering paradigms, comparing activation steering against prompt-level, score-space, and weight-space interventions, analyzing the interaction between the steering mechanism and the intervention site. Our new benchmark, supported by an extensive user study, demonstrates that localized activation steering establishes a new state-of-the-art in audio concept modulation.
title TADA! Tuning Audio Diffusion Models through Activation Steering
topic Sound
Machine Learning
url https://arxiv.org/abs/2602.11910