Bottleneck-localized steering beats full-network steering in music diffusion

measured in 1 paper

Staniszewski et al. use activation patching to localize a semantic bottleneck across three text-to-music diffusion models (AudioLDM2, Ace-Step, Stable Audio Open) [staniszewski-etal-2026-tada-tuning-audio-diffusion-models-through-activation-steering] Steering only at these 2-4 bottleneck layers, via mean-difference direction vectors or TopK SAE decoder columns, beats full-network steering [staniszewski-etal-2026-tada-tuning-audio-diffusion-models-through-activation-steering] On Ace-Step "Piano," LPAPS drops to 2.188 for layers {6,7} versus 2.707 for all 24 layers [staniszewski-etal-2026-tada-tuning-audio-diffusion-models-through-activation-steering] SAE-based bottleneck steering achieves the best preservation (LPAPS 1.973) [staniszewski-etal-2026-tada-tuning-audio-diffusion-models-through-activation-steering]

Structure

Linear Direction

Context

text-to-music, diffusion-models, activation-steering

Confirmed in models

AudioLDM2 · AudioLDM, Stable Audio Open · Stable Audio, Ace-Step · Ace-Step

Method

Activation patching, Difference-in-means direction, Sparse Autoencoders (SAE), Causal interventions (steering)

Papers

TADA! Tuning Audio Diffusion Models Through Activation Steering — Staniszewski, Łukasz, Zaleska, Katarzyna, Modrzejewski, Mateusz, Deja, Kamil2026 · arXiv:2602.11910