Mechanistic Interpretability of Text-to-MIDI Models: Probing, Lenses, and Steering
An arXiv paper (2608.06638) explores mechanistic interpretability methods applied to symbolic music generation, a field traditionally focused on audio model analysis. The research investigates two public text-to-MIDI systems with differing architectures: text2midi, designed as an encoder-decoder, and MIDI-LLM, based on a Llama 3.2 1B model enhanced with MIDI tokens. Through techniques like linear probing, logit and tuned lenses, activation patching, and difference-in-means steering, the authors uncover musically relevant structures and demonstrate how architecture influences their development and management. While both models allow for linear decoding of pitch, instrumentation, harmony, and texture, text2midi gradually refines predictions, contrasting with MIDI-LLM's abrupt transition into musical vocabulary. The paper, categorized as a cross-type submission on arXiv, suggests valuable insights for future symbolic music generation model design and interpretability.
Key facts
- Paper arXiv:2608.06638 analyzes mechanistic interpretability of text-to-MIDI models.
- Two models studied: text2midi (encoder-decoder) and MIDI-LLM (Llama 3.2 1B extended with MIDI tokens).
- Techniques used: linear probing, logit and tuned lenses, activation patching, difference-in-means steering.
- Pitch, instrumentation, harmony, and texture are linearly decodable in both models.
- text2midi refines predictions gradually across depth.
- MIDI-LLM works in textual basis before a sharp late rotation into musical vocabulary.
- Patching identifies late attenuation of prompt-driven instrument transfer.
- Steering produces bidirectional changes in register and polyphony in both systems.
- Announcement type: cross on arXiv.
- Symbolic music generation models are largely unexplored in mechanistic interpretability.
Entities
Institutions
- arXiv