- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies the application of circuit methods, used in recent LLM mechanistic interpretability research, to the YuE2 model - a generative audio model for music.
The authors analyze residual style vectors and the components that contribute most strongly to them, using circuit methods to compare activations between music generated with different style conditioning (and external classifier scores). The authors find that the resulting vectors depend on position.
The authors then propose computing position-agnostic activation vectors and component scores by weighting according to the KL divergence between next-token distributions with and without a style tag. This resembles CAA with KL-informed weighting, using circuit methods to identify responsible components (e.g., neurons and attention heads).
The authors find that their proposed techniques, e.g., manually modifying identified component activations, can be successfully applied to YuE2 to guide generation, using CLAP and Gemini scores to evaluate the outputs. The proposed techniques achieve higher style-hit counts and reliability than conventional CAA in this evaluation.
Critical strengths
Overall, I found this paper interesting. As the authors mention, mechanistic interpretability techniques have been underexplored in generative music. The authors identify a key problem in adapting existing techniques to music and provide evidence for this. The resulting solution is quite elegant. According to their extensive evaluations, the proposed approach (KL weighting) seems to work well in producing components and residual directions that can be used for style identification and modified to condition generation. This is also supported by the demo page. The analysis and additional ablations that the authors conduct are appreciated and useful. This work sets the scene for more mechanistic interpretability research into music models.
Critical weaknesses
Prefacing this with the fact that I’m not an expert in mechanistic interpretability, I have some questions about how significant the contributions of this paper are. I appreciate that this can be difficult to assess for papers that study applications and adaptations of techniques to the music domain. This paper builds heavily on existing techniques, and the primary methodological contribution outside of applying these techniques to a music model seems to be the introduction of KL weighting for component and residual-space differences. This is especially relevant because, as far as I can tell, Appendix D.1 shows that applying the same KL-weighting idea to CAA-style residual steering, instead of taking an ordinary mean, achieves reasonably close hit counts on this small evaluation and largely avoids the truncation issues mentioned in §4.5. While the proposed approach has slightly higher style hit counts, without uncertainty estimates, it is difficult to judge how robust these differences are. At the very least, this comparison and its implications for the claimed contribution should be foregrounded in the main paper. What exactly to take away from the paper could be made much clearer by discussing the most important ablations, especially those that separate the impacts of KL weighting, component selection, and intervention location. This would help establish how much additional benefit comes from the circuit method beyond the KL weighting itself.
Minor concerns:
l89: 'based-on' is a bit unclear. I think the proposal of the paper could be more clearly stated: e.g., we propose a technique for style discovery and control for generative music models by applying circuit control techniques, using YuE2 as a test bed.
l90: A bit pedantic, but perhaps 'state-of-the-art' here is a bit of a loaded descriptor, as generative music models exist in many forms and can be judged against many metrics and paradigms of interaction.
l164: Dynamic circuit steering is slightly confusing wording? As far as I can see the identified residual directions / activation components are static, applied equally at all positions, but calculated as a pooled vector weighted by KL divergence. The only dynamic aspect of this is the calculation of this vector (e.g., calculated as multi-position vs single-position as used in emotion circuits). The steering itself is applied statically.
l203–205: I find Eq. (1) difficult to parse. Section 3.1 defines as the “representation-analysis set,” but does not explicitly state what indexes. The inner sum also assumes that styles have been assigned integer indices without defining this convention.
I infer that denotes the set of lyrics used for representation analysis, denotes the set of five styles, and is the hidden state at layer and position for lyrics generated in style . With these definitions, the equation would read:
d^{(\ell,p)}_s \propto \sum_{i \in \mathcal{L}_{\mathrm{ext}}} \left( h^{(\ell)}_{i,s,p} - \frac{1}{|\mathcal{S}|} \sum_{s' \in \mathcal{S}} h^{(\ell)}_{i,s',p} \right), \qquad \left\\|d^{(\ell,p)}_s\right\\|_2 = 1.
Is this the intended interpretation? Please improve this.
l217–234: Minor, but I find the purpose of this ablation slightly confusing. My understanding is that you intervene on the identified component activations and then measure the change in the residual stream's dot-product projection onto the style directions. By construction, aren't the local effects largely expected? For example, ablating selected components gives
where is the component's write vector, and components are selected for large positive mean contributions to this projection.
I appreciate that Appendix B.2 also measures effects on the output distribution. Perhaps this should be analyzed in the main paper? It would help to clarify what the projection-based validation establishes beyond the direct local effects expected from the component-selection criterion.
Questions for authors
See weaknesses.
3 comments
Response to Reviewer fUFC (2/2): W3, minor points
W3. Ablations separating KL weighting, component selection and intervention location
Thank you for this suggestion. We ablated these factors separately, with 150 samples per condition; numbers are CLAP hits.
Position weighting and coverage (steering only the score stage; each representation uses its own selected components; same strength):
| Coverage | Equal weight | KL-weighted |
|---|---|---|
| First 128 steps | 63 | 80 |
| Whole segment | 70 | 77 |
Representation and intervention stage (each representation uses its own selected components; same strength within a stage):
| Representation | Score stage only | Music semantic token stage only | Both stages |
|---|---|---|---|
| KL-weighted (circuit) | 77 | 118 | 140 |
| Equal weight (first 128 steps) | 63 | 114 | 125 |
| Single-position | 34 | 60 | 65 |
| Random components | 32 | 68 | 77 |
| Blank | 30 | 30 | 30 |
Over the same positions, KL weighting is never worse than equal weighting (significantly better over the first 128 steps), whereas extending the coverage alone makes no significant difference; whether one or both stages are steered, the single-position representation is on par with random components, which shows that cross-position aggregation is necessary; with both stages steered, KL weighting is also significantly better than equal weighting. Each stage works on its own, and combining both works best. The same KL-weighted difference applied to all components is clearly weaker than the circuit, and applied to random components has no effect. For the number and type of components, see our response to 4oTW (W4) (halving or doubling the circuit changes little; control comes mainly from the attention heads); for components vs. a single-layer residual stream, see W2.
Minor points
- l89: We adopt your suggested wording: "We propose a technique for discovering and controlling style in generative music models via circuit methods, using YuE2 as a testbed."
- l90: Agreed; an unqualified "state-of-the-art" is indeed loaded. Among open-source lyrics-to-song models, YuE2 is at the leading level; the revised version will state the scope, e.g., "one of the strongest open-source lyrics-to-song models", and cite the comparisons in the YuE2 technical report.
- l164: Your understanding is correct: the steering itself is applied statically. "Dynamic" refers to the source of the direction: the style representation varies with position in both the score and the music semantic tokens, and is aggregated across positions with KL weighting, which is one of the main contributions of the paper (W1, W3). The revised version may consider adjusting the wording, e.g., "circuit steering with KL-weighted dynamic representations".
- l203–205: Your reading matches our intended meaning, and we adopt it directly: denotes lyrics in the representation-analysis set , is the set of styles, and the inner term is written as .
- l217–234: Agreed; the projection-based ablation mainly tests what a component writes in its own layer, which is expected given the selection criterion. In the revised version, the causal validation in the main text will be based mainly on the output distribution (already reported in §4.4): zeroing only 0.23% of the MLP neurons moves the output away from the style-tag output by about 80% of the distance reached when the tag is removed, and enhancing only a few attention heads recovers most of the output difference; norm-matched random components have almost no effect, and only zeroing attention heads harms the output indiscriminately. The projection readout will be moved to the appendix.
Questions to the authors: please see our responses to W1–W3.
Thank you again for your careful review and for all the questions you raised. We hope that the additional experiments and clarifications have helped address your concerns, and we would be grateful if you could take them into account when revisiting your evaluation. We would be glad to answer any further questions.
Thank you again for your time and thoughtful review.
Thank you again for your time and thoughtful review. As the discussion period is drawing to a close, we would greatly appreciate the opportunity to further clarify and improve our work before the deadline.
Could you kindly let us know if our previous response, adequately addresses your concern? We look forward to your feedback and would be more than happy to address any remaining questions you may have.
Response to Reviewer fUFC (1/2): W1, W2
W1. Significance of the contribution
Thank you for your positive assessment and candid comments. We stand by the view that the contribution of this paper is not KL weighting alone, but four parts: the stage and position analysis, the KL-weighted cross-position aggregation derived from it, the circuit analysis showing which components write this representation into the residual stream, and the generality of the method across models and data (General Response, point 1). Unlike single-position approaches in NLP, music generation models are mostly multi-stage, and in YuE2 a circuit assembled from the single-position representation is on par with random components; KL weighting places the weight where the tag actually changes the output and roughly doubles the hits of the single-position one (ablation tables under W3). The revised version will present the four parts side by side as contributions.
W2. KL-weighted residual steering is already close to the circuit; no uncertainty estimates
Thank you for pointing this out; the revised version will move this comparison and its implications for the contribution into the main text. Appendix D.1 has only 70 samples and no uncertainty estimates, so we repeated the comparison on new test lyrics: both methods use the same KL-weighted representation, the circuit injects it into the localized components and residual steering adds it to the residual stream at a single layer, with the strengths taken from D.1; both steer only the score stage. Each method has 75 samples, with seeds not involved in any selection.
In the difference row, "a vs. b" counts the pairs where only the circuit or only residual steering hits (exact McNemar test). The circuit has more hits under all three judges (significantly under Music Flamingo) and a significantly higher margin; across four shared strength levels, it is significantly better at two and comparable at the other two.
The revised version will therefore state the contribution as follows: the KL-weighted stage representation is one of the main contributions, and component localization still adds a gain on top of it. Injecting the same representation into the localized components yields more hits at the same content cost, whereas random components in the same layers have no effect. The circuit also requires no layer selection, can control each stage independently, and preserves the melody better (General Response, point 2).