There is a quiet problem at the heart of every “steerable” music model: the thing you are offered as a control may not be the thing that moves the music.
MI-MIDI 2 was posted to arXiv on 23 October 2026 (arXiv:2610.32007) by Jakub Poćwiardowski and Mateusz Modrzejewski. The Comments field gives 20 pages, 6 figures, 24 tables and an atlas of audio examples, and names no venue, so this is an unreviewed preprint. The demo atlas is live at jpocwiar.github.io/MI-MIDIv2-Demo/, which we checked.
The borrowed warning
Sparse autoencoders are the current workhorse of mechanistic interpretability. Train one on a model’s internal activations and it decomposes them into a large set of sparse features, each of which you can then try to characterise — without anyone having decided in advance what the categories should be.
The paper opens with the problem that the language-model literature has already run into, and states it precisely:
in language models the evidence used to interpret these features has proven an unreliable guide to which of them steer the output
That sentence is the whole motivation. You find a feature. You look at what activates it, write a description, and the description is persuasive. Then you intervene on that feature and the output does not change in the way the description predicts. The interpretation was a story about correlations; control is a causal claim, and the first does not establish the second.
And then the gap the paper is filling:
Whether interpretations of SAE features predict control in music generators has not, to our knowledge, been tested.
The protocol, and why the controls are the contribution
The authors build an end-to-end protocol for discovering and evaluating SAE features in text-to-MIDI generation, applied to two models: text2midi and MIDI-LLM.
Features are characterised three ways, and each kind of evidence gets its own control — which is the part that distinguishes this from a feature-atlas paper:
| Evidence used to interpret a feature | Its control |
|---|---|
| Correlations with musical metrics | note-density adjustment |
| Shared instrument programs | matched inactive excerpts |
| Automated descriptions | permutation calibration |
Each control neutralises the obvious confound. Correlations with musical metrics are vulnerable to note density driving everything, so density is adjusted for. A feature that looks instrument-specific may simply fire on busy passages, so it is compared against matched excerpts where it is inactive. And automated descriptions will always sound plausible, so they are calibrated against permutations — descriptions generated for shuffled data, which establishes the floor that a meaningless description scores.
That last one deserves emphasis. A permutation control on automated descriptions is a way of asking how impressive a description has to be before it means anything. Without it, every feature has a convincing label.
Then the causal test. Candidates are tested by intervention, measured directly in the generated MIDI, against two baselines: unmodified generation and a matched random direction. Candidates are ranked in screens, and survivors are confirmed on new prompts.
The random-direction baseline is the one that catches the most common self-deception in steering work. Push any direction in a representation hard enough and the output changes. The question is whether this direction changes it more, or more specifically, than an arbitrary one.
What survived
The abstract reports two findings, and both are narrow in a way that reads as honesty rather than weakness.
Descriptions beat permutation controls for instrumentation. So automated descriptions do carry real signal — about which instruments a feature relates to, at least — above the level a shuffled description achieves.
A feature whose description names a synth-lead program is singled out as an example that carried through to intervention. The abstract cuts off mid-sentence in the arXiv record, so we are not going to extrapolate the rest.
The framing the paper offers for the whole exercise is the sentence to remember: “Evidence that survives these controls points to features that steer the output.” Not evidence that is convincing. Evidence that survives.
Why this matters for anyone making music with these tools
The practical consequence is about what to trust in a product.
Interpretability features are starting to appear in creative tooling as sliders — “more jazzy”, “sparser”, “brighter” — derived from exactly this kind of analysis. This paper is a reminder that the pipeline from “we found a feature that correlates with X” to “here is a slider for X” has a step in it that is easy to skip, and that skipping it produces controls which demo well and fail in use.
If you are building on an SAE-derived control, the testable question is the one here: does intervening on it beat a random direction of the same magnitude, measured in the output rather than in the latent space, on prompts you did not use to find the feature?
Two caveats on scope. This is MIDI, not audio — symbolic note data, where “instrument program” is a discrete field you can read directly, which is precisely what makes the instrumentation result measurable. Whether the same protocol transfers to audio models, where nothing is labelled in the output, is a different and harder question. And it is an unreviewed preprint on two text-to-MIDI models.
Sources
- MI-MIDI 2: From Observation to Control with Sparse Autoencoders in Text-to-MIDI Generation — arXiv:2610.32007
- MI-MIDI v2 demo atlas with audio examples
- arXiv PDF
- arXiv listing for cs.SD