AR&D: A Framework for Retrieving and Describing Concepts for Interpreting AudioLLMs
Authors: Townim Faisal Chowdhury, Ta Duc Huy, Siqi Pan, Jeremy Stoddard, Zhibin Liao
Organizations: Australian Institute for Machine Learning, University of Adelaide · Dolby Laboratories · School of Computer and Mathematical Sciences, University of Adelaide, Australia
Despite strong performance in audio perception tasks, large audio-language models (AudioLLMs) remain opaque to interpretation. A major factor behind this lack of interpretability is that individual neurons in these models frequently activate in response to several unrelated concepts. We introduce the first mechanistic interpretability framework for AudioLLMs, leveraging sparse autoencoders (SAEs) to disentangle polysemantic activations into monosemantic features. Our pipeline identifies representative audio clips, assigns meaningful names via automated captioning, and validates concepts through human evaluation and steering. Experiments show that AudioLLMs encode structured and interpretable features, enhancing transparency and control. This work provides a foundation for trustworthy deployment in high-stakes domains and enables future extensions to larger models, multilingual audio, and more fine-grained paralinguistic features. Project URL: https://townim-faisal.github.io/AutoInterpret-AudioLLM/
Figures & tables
Figure 1: Overview of the Audio Retrieve and Describe (AR&D) pipeline to discover and naming interpretable concepts in AudioLLM. Stage 1 : The SAE is trained to reconstruct representations x from the AudioLLM, yielding a latent space of sparse, monosemantic features. Stage 2 : Using a probing dataset A , we compute SAE activations Z and calculate a representativeness scores by F(⋅) for each feature, selecting the p most and least representative audio clips ( Hk and Lk ) per feature. Stage 3 : We filter top features using monosemanticity scores derived from Hk and Lk , and interpret them by generating and summarizing captions from representative clips Hk , producing a final set of human-understandable concepts.
Figure 2: Illustration of the steering mechanism. From a given layer of AudioLLM, the obtained input x is transformed into an SAE representation z . The k -th feature ( i.e., the targeted concept) is then replaced with a predefined steering value ( e.g., up from 2.5 to 4.0 in the example). The modified representation z^ is processed by a TopK operator and decoded into x^ (Eq. 1 ), which is subsequently fed through the rest AudioLLM layers (as marked by the dotted lines), replacing x , allowing fine-grained control over specific features in the model.
Method
MS ↑
Precision ↑
Recall ↑
F1 ↑
mAP ↑
Poly. Feats. [ 4 ]
1.14
0.05
0.20
0.08
0.05
Rand. Rep.
2.13 ±0.08
0.11 ±0.05
0.23 ±0.06
0.15 ±0.04
0.10 ±0.05
Mean Act. [ 26 ]
4.76
0.42
0.42
0.42
0.36
Coverage
5.28
0.46
0.44
0.45
0.39
AR&D ( l =5)
9.17
0.48
0.51
0.49
0.47
AR&D ( l =16)
9.31
0.52
0.58
0.55
0.53
Table 1: Evaluation of interpretability on the FSD50k dataset, showing highest monosemanticity scores (MS), precision, recall, F1, and mean Average Precision (mAP). Higher values indicate better semantic alignment and feature interpretability.
Method
Expert-rated score ↑
CLAP Cosine ↑
BERTScore↑
Poly. Feats.
2.13 ±0.79
0.47 ±0.41
0.23 ±0.41
AR&D
4.29 ±0.81
0.84 ±0.20
0.92 ±0.11
Table 2: Human and semantic similarity evaluations.
Method
Neutral → Happy
Sad → Happy
Happy → Angry
Female → Male
Poly. Feats.
0.13
0.08
0.04
0.09
AR&D
0.75
0.68
0.47
0.61
Table 3: Evaluation of feature steering sensitivity on targeted emotion and gender concepts. Higher scores indicate a stronger effect of the intervention on the model’s output.
Method
Expansion
Layer
Factor
5
16
26
Poly. Feats.
-
0.68
1.02
1.14
AR&D
4
6.86
7.34
9.17
8
7.01
7.85
9.31
16
7.06
7.89
9.33
Table 4: Monosemanticity score on different layer and expnsion factor. Higher is better.
Multimodal model interpretability studies how AI systems combine and use images, text, audio, video, and other signals, and how to explain their decisions.