ICLR 2026

Discovering and Steering Interpretable Concepts in Large Generative Music Models

Nikhil Singh*Dartmouth College

Manuel Cherep*MIT Media Lab, Massachusetts Institute of Technology

Pattie MaesMIT Media Lab, Massachusetts Institute of Technology

* Equal Contribution

Open Interactive Example Dashboard Paper Code

Abstract

The fidelity with which neural networks can now generate content such as music presents a scientific opportunity: these systems appear to have learned implicit theories of such content's structure through statistical learning alone. This offers a potentially new lens on theories of human-generated media. When internal representations align with traditional constructs (e.g. chord progressions in music), they show how such categories can emerge from statistical regularities; when they diverge, they expose limits of existing frameworks and patterns we may have overlooked but that nonetheless carry explanatory power. In this paper, focusing on autoregressive music generators, we introduce a method for discovering interpretable concepts using sparse autoencoders (SAEs), extracting interpretable features from the residual stream of a transformer model. We make this approach scalable and evaluable using automated labeling and validation pipelines. Our results reveal both familiar musical concepts and coherent but uncodified patterns lacking clear counterparts in theory or language. As an extension, we show such concepts can be used to steer model generations. Beyond improving model transparency, our work provides an empirical tool for uncovering organizing principles that have eluded traditional methods of analysis and synthesis.

Examples of Steering Outputs

From left to right: original output, steered output, representative examples of target concept used for steering.

Steering Set 1

Baseline

Steered Output

Reference 1

Reference 2

Reference 3

Steering Set 2

Baseline

Steered Output

Reference 1

Reference 2

Reference 3

Steering Set 3

Baseline

Steered Output

Reference 1

Reference 2

Reference 3

Prompting Comparison

Comparison of prompting effectiveness vs. discovered features for concept “Silence”.

Prompting

Prompt 1

Prompt 2

Prompt 3

Discovered Feature

Example 1

Example 2

Example 3

Concept Examples

Examples of clips that highly activate discovered concepts (manually labeled; see dashboard for automatically labeled examples).

Bb4 Pedal Tone

Example 1

Example 2

Example 3

Acoustic Folk Singer Songwriter

Example 1

Example 2

Example 3

Baroque Harpsichord

Example 1

Example 2

Example 3

Electronic Beeps and Boops

Example 1

Example 2

Example 3

Classic Rock

Example 1

Example 2

Example 3

Clipping

Example 1

Example 2

Example 3

Drum Roll

Example 1

Example 2

Example 3

Epic Battle Music

Example 1

Example 2

Example 3

Eurodance

Example 1

Example 2

Example 3

Guitar Solo

Example 1

Example 2

Example 3

Hardstyle

Example 1

Example 2

Example 3

Midi Piano

Example 1

Example 2

Example 3

Midtempo Jazz

Example 1

Example 2

Example 3

Music and Rain

Example 1

Example 2

Example 3

Music and Waves

Example 1

Example 2

Example 3

New Age Synth Hi-Hat Groove

Example 1

Example 2

Example 3

Non-Western Percussion

Example 1

Example 2

Example 3

One Instrument One Note

Example 1

Example 2

Example 3

Oscillating Bell

Example 1

Example 2

Example 3

Quiet Silence

Example 1

Example 2

Example 3

Saxophone

Example 1

Example 2

Example 3

Ska

Example 1

Example 2

Example 3

Taiko

Example 1

Example 2

Example 3