Katarzyna Zaleska

N(𝐡¯c(i)−𝐡¯c~(i)),\mathbf{v}_{c}^{\text{CAA}}=\frac{\mathbf{v}_{c}}{\|\mathbf{v}_{c}\|_{2}}\text{, where}\quad\mathbf{v}_{c}=\frac{1}{N}\sum_{i=1}^{N}\left(\bar{\mathbf{h}}_{c}^{(i)}-\bar{\mathbf{h}}_{\tilde{c}}^{(i)}\right), (2) with 𝐡¯\bar{\mathbf{h}} denoting cross-attention outputs averaged across temporal frames. During generation, we steer by modifying cross-attention outputs at functional layers as 𝐡l′=ReNorm​(𝐡l+α⋅𝐯cCAA,𝐡l)\mathbf{h}^{\prime}_{l}=\text{ReNorm}(\mathbf{h}_{l}+\alpha\cdot\mathbf{v}_{c}^{\text{CAA}},\mathbf{h}_{l}), where α∈ℝ\alpha\in\mathbb{R} controls steering strength, with positive values adding and negative values removing the concept. ReNorm is a re-normalization operation, ensuring that the norm of the output activations 𝐡l′\mathbf{h}^{\prime}_{l} matches the one before intervention 𝐡l\mathbf{h}_{l}: ReNorm​(𝐡l′,𝐡l)=𝐡l′‖𝐡l′‖2⋅‖𝐡l‖2.\text{ReNorm}(\mathbf{h}^{\prime}_{l},\mathbf{h}_{l})=\frac{\mathbf{h}^{\prime}_{l}}{\|\mathbf{h}^{\prime}_{l}\|_{2}}\cdot\|\mathbf{h}_{l}\|_{2}. (3) Sparse Autoencoders (SAEs). To further discover interpretable features within cross-attention activations, we train a TopK SAE (Gao et al., 2025; Bussmann et al., 2024) on the functional layer with the highest response during activation patching. The SAE with encoder weights 𝐖enc∈ℝm​d×d\mathbf{W}_{\text{enc}}\in\mathbb{R}^{md\times d}, decoder weights 𝐖dec∈ℝd×m​d\mathbf{W}_{\text{dec}}\in\mathbb{R}^{d\times md}, and bias 𝐛pre∈ℝd\mathbf{b}_{\text{pre}}\in\mathbb{R}^{d} maps activations 𝐡∈ℝd\mathbf{h}\in\mathbb{R}^{d} to a sparse code 𝐟∈ℝm​d\mathbf{f}\in\mathbb{R}^{md} (with mm being the expansion factor): 𝐟=TopK​(𝐖enc​(𝐡−𝐛pre)),\mathbf{f}=\text{TopK}\left(\mathbf{W}_{\text{enc}}(\mathbf{h}-\mathbf{b}_{\text{pre}})\right), (4) and further reconstructs them: 𝐡^=𝐖dec​𝐟+𝐛pre.\hat{\mathbf{h}}=\mathbf{W}_{\text{dec}}\mathbf{f}+\mathbf{b}_{\text{pre}}. (5) The TopK​(⋅)\text{TopK}(\cdot) operation retains only the kk largest activations in the autoencoder’s latent space, zeroing out the rest. The SAE is trained to minimize reconstruction error ‖𝐡−𝐡^‖22\|\mathbf{h}-\hat{\mathbf{h}}\|_{2}^{2}. To identify concept-specific features, we use two contrastive sets of NN prompts (𝒫c,𝒫c~)(\mathcal{P}_{c},\mathcal{P}_{\tilde{c}}) for concept cc to compute importance scores using a TF-IDF-based criterion: score​(j,c)=μj​(𝒫c)⏟TF⋅log⁡(1+1μj​(𝒫c~)+ϵ)⏟IDF,\text{score}(j,c)=\underbrace{\mu_{j}(\mathcal{P}_{c})}_{\text{TF}}\cdot\underbrace{\log\left(1+\frac{1}{\mu_{j}(\mathcal{P}_{\tilde{c}})+\epsilon}\right)}_{\text{IDF}}, (6) where μj​(𝒫𝒸)=1|𝒫𝒸|​∑𝐡∈𝒫𝒸fj​(𝐡)\mu_{j}(\mathcal{P_{c}})=\frac{1}{|\mathcal{P_{c}}|}\sum_{\mathbf{h}\in\mathcal{P_{c}}}f_{j}(\mathbf{h}) is the mean activation of feature jj in the SAE’s latent space on generated audio with prompts 𝒫𝒸\mathcal{P_{c}}. Features that activate highly for concept cc but rarely for other samples receive high scores. We select the top-τc\tau_{c} scoring features ℱc\mathcal{F}_{c} and, by summing their corresponding decoder columns, we construct the steering vector 𝐯cSAE\mathbf{v}_{c}^{\text{SAE}} as 𝐯cSAE=∑j∈ℱc𝐖dec​[:,j],\mathbf{v}_{c}^{\text{SAE}}=\sum_{j\in\mathcal{F}_{c}}\mathbf{W}_{\text{dec}}[:,j], (7) and add it directly to the output of the cross-attention layer: 𝐡l′=𝐡l+α⋅𝐯cSAE.\mathbf{h}^{\prime}_{l}=\mathbf{h}_{l}+\alpha\cdot\mathbf{v}_{c}^{\text{SAE}}. (8) 4 Experiments 4.1 Layer localization To measure the importance of individual cross-attention layers in text-to-audio models, we construct a dataset of counterfactual prompt pairs. We consider the following musical concepts: vocal gender (female vs. male), tempo (slow vs. fast), mood (happy vs. sad), and categories such as instruments (drums, flute, guitar, maracas, trumpet, violin), and genres (jazz, techno, reggae). For each contrasting pair (c,c~)(c,\tilde{c}), we select captions from the MusicCaps (Agostinelli et al., 2023) dataset that contain keywords associated with concept cc (e.g., ‘female voice’) but do not contain any terms related to the alternative variant c~\tilde{c} (e.g., ‘male voice’). For genres and instruments, we replace concepts (e.g., ‘violin’) with alternatives (e.g., ‘trumpet’). We select up to 256 such prompts as target ones 𝒫c\mathcal{P}_{c} and use GPT-4 (Achiam et al., 2023) to generate alternatives 𝒫c~\mathcal{P}_{\tilde{c}} by replacing concept-associated terms with their counterparts while preserving all other content. Examples of prompt pairs and concept replacements are provided in the App. A. We apply the activation patching procedure described in Section 3 to three state-of-the-art audio diffusion models: AudioLDM2 (Liu et al., 2024), Stable Audio Open (Evans et al., 2025), and Ace-Step (Gong et al., 2025). We generate waveforms of 10 seconds (AudioLDM2, Stable Audio Open) and 30 seconds (Ace-Step), using 8 different random seeds per prompt, resulting in 2048 generations per concept. The impact of layer ll on concept cc is calculated as: Impact​(l,c)=sim​(l←c,l′←c~)−sim​(l←c~,l′←c~)sim​(l←c,l′←c)−sim​(l←c~,l′←c~),\text{Impact}(l,c)=\frac{\text{sim}(l\leftarrow c,l^{\prime}\leftarrow\tilde{c})-\text{sim}(l\leftarrow\tilde{c},l^{\prime}\leftarrow\tilde{c})}{\text{sim}(l\leftarrow c,l^{\prime}\leftarrow c)-\text{sim}(l\leftarrow\tilde{c},l^{\prime}\leftarrow\tilde{c})}, (9) where sim​(l←c1,l′←c2)\text{sim}(l\leftarrow c_{1},l^{\prime}\leftarrow c_{2}) denotes the audio-text similarity between the concept name and a generation where layer ll inputs prompt c1c_{1} while all other layers l′=ℒ∖{l}l^{\prime}=\mathcal{L}\setminus\{l\} receive c2c_{2}. We use MuQ (Zhu et al., 2025) for assessing mood, tempo, instruments, and genres, and CLAP (Wu et al., 2022) for distinguishing vocal gender. Fig. 3 presents the layer-wise impact scores for each model. Across all three architectures, we observe that a small subset of layers concentrates the control over musical concepts. In AudioLDM2, with the U-Net architecture, we localize key components in the decoder, specifically the layers {44,45,50,51} (4 out of 64 cross-attentions), with a slight contribution from the layers in-between (46–49). In transformer-based architectures, we observe an intense concentration of control in the middle layers (2 out of 24). Namely, in Ace-Step, cross-attentions {6,7} exhibit strong influence across all concept categories, suggesting their role as a semantic bottleneck. In Stable Audio Open, similar concentration arises in layers {11,12}. These findings suggest that audio diffusion models develop interpretable, functionally specialized layers that are shared across various audio concepts. Additionally, our experiments indicate that such specialization phenomenon is not limited to a single model, but is a general property of text-to-music DMs. Figure 3: Functional cross-attention layers in AudioLDM2 (Liu et al., 2024), Stable Audio Open (Evans et al., 2025), and ACE-Step (Gong et al., 2025) models. We demonstrate that singular layers control different musical concepts, including vocal gender, tempo, mood, instruments, and genres across diverse audio diffusion architectures. 4.2 Steering Audio Diffusion Models Given a strong specialization of cross-attention layers, we further evaluate their usefulness on a downstream task of concept steering with audio DMs. The goal is to modulate the generated audio to increase the likelihood of including the concept cc according to the steering strength α\alpha, while preserving other audio characteristics untouched. Evaluation Metrics. We evaluate steering across four dimensions, over a range of steering strengths α∈{αmin,…,αmax}\alpha\in\{\alpha_{\text{min}},\ldots,\alpha_{\text{max}}\}. Preservation measures how well the original audio characteristics are maintained with steering using LPAPS (Iashin and Rahtu, 2021) and FAD (Kilgour et al., 2019) computed against the unsteered baseline (α=0\alpha=0), averaging metrics over all steering strengths α\alpha. Δ\Delta Alignment quantifies steering effectiveness as the difference in audio-text similarity between the maximum and minimum steering strengths, measured using CLAP (Wu et al., 2022) and MuQ (Zhu et al., 2025) models; higher values indicate a higher degree of concept manipulation, and thus higher steering expressiveness. Smoothness captures the consistency of transitions, computed as the standard deviation of consecutive alignment differences across α\alpha, where lower values indicate smoother interpolation. Finally, Audio Quality is assessed using Audiobox Aesthetics (Tjandra et al., 2025), scoring average Content Enjoyment (CE), Content Usefulness (CU), Production Complexity (PC), and Production Quality (PQ). Details. We evaluate steering methods with Ace-Step, generating 30-second audio clips with 60 diffusion steps. For Contrastive Activation Addition (CAA), we compute steering vectors leveraging contrastive prompt pairs (see App. B for examples) and apply them with uniform strengths α∈{−100,−90,…,90,100}\alpha\in\{-100,-90,\ldots,90,100\}. Evaluation is performed on 100100 diverse prompts spanning diverse music styles and concepts, allowing us to assess the steering effect on wide range of generations. We train SAE on layer {7}\{7\} activations from generations from the MusicCaps (Agostinelli et al., 2023) prompts. We conduct a hyperparameter search over expansion factors (m∈{2,4,8,16}m\in\{2,4,8,16\}) and sparsity parameters (k∈{16,32,64}k\in\{16,32,64\}). We select the best configuration based on the lowest reconstruction error and fraction of both dead and high-frequency features, which is SAE (m=4,k=64)m=4,k=64) trained for 1515 epochs. Concept-specific features are selected based on generations from contrastive prompt pairs (same as in the case of CAA), using TF-IDF scoring (Eq. 6) with the following top-τ\tau values: τ=20\tau=20 for piano, τ=20\tau=20 for vocal gender, τ=20\tau=20 for tempo, and τ=40\tau=40 for mood. Results. Table 1 presents the qualitative results of our steering experiments on the Ace-Step model, comparing our targeted intervention strategy against global baselines. We compare steering exclusively the identified functional layers ({6,7}\{6,7\}) against steering all layers (ℒ\mathcal{L}) and, crucially, an ablation setting where we steer all layers except the functional ones (ℒ∖{6,7}\mathcal{L}\setminus\{6,7\}). The results provide striking validation of our localization hypothesis. We observe a ”semantic bottleneck”, consisting of controlling layers, within which the model’s steering capacity for high-level concepts is almost entirely concentrated. As shown in Table 1, steering these two layers alone ({6,7}\{6,7\}) yields high alignment scores across all concepts. Conversely, when we apply steering to the remaining 22 layers while leaving the functional layers untouched (ℒ∖{6,7}\mathcal{L}\setminus\{6,7\}), the ability to control the generation collapses. Beyond successful alignment, targeted steering demonstrates superior preservation of the original audio’s characteristics. By limiting the intervention to the semantic bottleneck, we minimize collateral damage to unrelated acoustic features. This is evidenced by the Preservation metrics (LPAPS and FAD), where our steering consistently maintains lower distances to the original audio compared to using the vast majority of non-functional layers. Furthermore, while global steering (ℒ\mathcal{L}) often degrades the overall fidelity of the output, our targeted approach maintains Audio Quality scores (CE, CU, PC, PQ) closer to the original, unsteered generations, showing less degradation or even improvement in production value or listening experience. Finally, the results using Sparse Autoencoders (SAE​({7})\text{SAE}(\{7\})) further corroborate these findings. By steering along specific feature directions within a single functional layer (Layer 7), we achieve alignment scores competitive with, and occasionally surpassing, the raw activation steering of combined layers, offering the highest degree of interpretability. This is a noteworthy case, given the recent studies by Kantamneni et al. (2025) questioning the practical utility of SAEs compared to standard baselines in LLMs. Concept Layers (#) Preservation (↓\downarrow) Δ\Delta Alignment (↑\uparrow) Smoothness (↓\downarrow) Audio Quality LPAPS FAD CLAP MuQ CLAP MuQ CE↑\uparrow CU↑\uparrow PC↑\uparrow PQ↑\uparrow No steering (α=0\alpha=0) - - - - - - 6.796.79 7.627.62 5.515.51 7.407.40 Piano ℒ\mathcal{L} 24 2.7072.707 0.4860.486 0.1060.106 0.1740.174 0.0040.004 0.0040.004 6.616.61 7.507.50 5.445.44 7.297.29 ℒ∖{6,7}\mathcal{L}\setminus\{6,7\} 22 2.5242.524 0.4670.467 −0.011-0.011 −0.016-0.016 0.0040.004 0.0020.002 6.596.59 7.477.47 5.455.45 7.287.28 {6,7}\{6,7\} 2 2.1882.188 0.2760.276 0.1000.100 0.1900.190 0.0040.004 0.0050.005 6.646.64 7.527.52 5.475.47 7.307.30 SAE({7}\{7\}) 1 1.9731.973 0.2620.262 0.1120.112 0.1910.191 0.0020.002 0.0020.002 6.796.79 7.637.63 5.225.22 7.407.40 Vocal Gender