acceptodds
Under review as a conference paper at ICLR 2027

Music Circuit: Trace and Control Style Dynamics in Large Music Generation Model

Abstract

Musical style, a global attribute that functions throughout a whole piece, is a lever widely used by users in modern large music generation models. Yet how style is represented inside these models, and which components drive it, remains largely unexplored. In language models, circuit analysis traces a representation to a sparse sub-network of a few key components, namely attention heads and MLP neurons, and has been used successfully to explain and control global attributes such as emotion. In this work, we use circuit analysis to study the mechanisms behind musical style. Based on YuE2, we build a dataset that pairs the same lyrics with five styles and design a three-step framework. First, stage localization shows that the style forms mainly in the score-planning stage. Second, static observation across the positions of this stage further reveals that style is not a fixed vector but a dynamic representation across tokens. Third, we therefore merge the activation differences across positions into one direction per style, and trace the neurons and attention heads that drive this direction into a sparse circuit. Without any style text, steering this circuit successfully transfers one piece into the target style, matching and even exceeding the style prompt. To our knowledge, this is the first work to analyze and control musical style at the circuit level in a large music generation model. Audio and score examples are available at https://aristo233.github.io/Music_Circuit/.

Primary area:
interpretability and explainable AI
Keywords:
Music AI, Mechanistic Interpretability, Circuit Discovery

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_fUFCOpenReview ↗
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 IextI_{\mathrm{ext}} as the “representation-analysis set,” but does not explicitly state what ii indexes. The inner sum also assumes that styles have been assigned integer indices without defining this convention.

I infer that Lext\mathcal{L}_{\mathrm{ext}} denotes the set of lyrics used for representation analysis, S\mathcal{S} denotes the set of five styles, and hi,s,p(ℓ)h^{(\ell)}_{i,s,p} is the hidden state at layer ℓ\ell and position pp for lyrics ii generated in style ss. 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

ΔSlocal=−∑j∈Jaj⟨wj,ds⟩,\Delta S_{\mathrm{local}} = -\sum_{j\in J} a_j\langle w_j,d_s\rangle,

where wjw_j 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
AuthorsOpenReview ↗

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.

MethodCLAP hitsMuQ-MuLan hitsMusic Flamingo hitsTruncationCLAP margin
Circuit4236462−0.020
KL-weighted residual steering3429360−0.055
Difference (circuit − residual)14 vs. 6, p = 0.1211 vs. 4, p = 0.1214 vs. 4, p = 0.031—+0.035 (+0.007, +0.055)

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).

AuthorsOpenReview ↗

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):

CoverageEqual weightKL-weighted
First 128 steps6380
Whole segment7077

Representation and intervention stage (each representation uses its own selected components; same strength within a stage):

RepresentationScore stage onlyMusic semantic token stage onlyBoth stages
KL-weighted (circuit)77118140
Equal weight (first 128 steps)63114125
Single-position346065
Random components326877
Blank303030

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: ii denotes lyrics in the representation-analysis set Iext\mathcal I_{\mathrm{ext}}, S\mathcal S is the set of styles, and the inner term is written as 1∣S∣∑s′∈S\frac{1}{|\mathcal S|}\sum_{s'\in\mathcal S}.
  • 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.

AuthorsOpenReview ↗

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.

Reviewer_dHzMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies where musical style is formed in YuE2. It uses KL-weighted style directions to select neurons and attention heads, then controls generation without style text. The evaluation includes automatic scores, random controls, and a preliminary listening study.

Critical strengths

The ablation, enhancement, and norm-matched random controls provide stronger evidence than activation correlation alone. The experiments also show better generation stability than CAA.

Critical weaknesses

The advantage over text prompting is mainly supported by automatic scores, while both CLAP and Gemini were used in data screening. In the preliminary non-pop listening questions, the prompt was chosen as closest to the target style in 27/43 cases, compared with 14/43 for the circuit. The difference was not significant, so better perceptual performance is not yet conducted. (Appendices A and D.2–D.3) Also, the style experiments generate new scores. The melody-similarity analysis does not perform style transfer with a fixed melody. (Appendix D.5)

Questions for authors

Does “better than prompting” refer only to automatic evaluation? What evidence supports this claim independently of the screening evaluators? Do you have style-control results with a fixed score?

3 comments
AuthorsOpenReview ↗

Response to Reviewer dHzM

W. Automatic evaluation, human evaluation and fixed melody

Thank you for your comments. We respond in three points.

(1) Metrics independent of the filtering. We added two judges that were not involved in data filtering: the circuit exceeds the full prompt under MuQ-MuLan and matches it under Music Flamingo, consistent with the paper's claim that the circuit matches or even exceeds the style prompt (table and protocol in our response to 4oTW (W2)). Beyond binary hits, the circuit's style margin is also significantly higher than that of the full prompt under both judges (see our response to Us9W (W4)). In the score-preserving setting, which the style prompt cannot control, and on LeVo 2, all three judges show the circuit clearly stronger.

(2) Subjective vs. objective evaluation. The paper claims that the circuit matches or even exceeds the style prompt, not that it is perceptually better, and the human evaluation is consistent with this: both are clearly above the empty-tag baseline, and the circuit matches the style prompt. The revised version will state that "exceeds" holds only under CLAP and MuQ-MuLan; for the reasons behind the discrepancy, see General Response, point 4.

(3) Style transfer with a fixed score. Appendix D.5 concerns the main experiment, which rewrites the score. We added results with a fixed score: steering only the music semantic token stage, the circuit's style hits are more than twice those of the tag alone at the same stage; compared directly with the score under the YuE2 technical report's protocol, the circuit preserves most of the melody, with the loss concentrated in metal rock, whereas single-layer residual steering loses much more of it. Results are in General Response, point 2. Demo: Score-Preserving Style Transfer.

Q. Follow-up questions

  • Does the claim refer only to automatic evaluation? Yes, and only to CLAP and MuQ-MuLan; the revised version will rewrite the abstract and §4.5 accordingly.
  • Evidence independent of the filtering judge: under MuQ-MuLan, both the hits and the style margin of the circuit are higher than those of the full prompt; under Music Flamingo, the circuit matches the style prompt when generating from lyrics and is clearly higher with a fixed score and on LeVo 2. See W, point (1).
  • Style control with a fixed score: yes, and this is where the circuit's advantage is clearest (118 / 95 vs. 51 / 44; 133 vs. 53 under Music Flamingo), with most of the melody preserved. See W, point (3). Demo: Score-Preserving Style Transfer.

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.

AuthorsOpenReview ↗

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.

Thank you for the detailed response and the additional work. The new judges, which were not used for data filtering, and the fixed-score results solve some of my concerns well. Moreover, I appreciate that the claim of better performance will be limited to the automatic evaluation in the revised version. But here is my point: with 11 listeners and 27/43 versus 14/43 picks, I think the listening study cannot separate the circuit from the prompt, rather than showing that they match. So the revision part cannot support me to raise my score further. Thank you again for the clear clarifications. I guess more solid listening tests will improve this work largely. Overall this is a valuable work. I especially like the idea that use circuit analysis to understand music style mechanism ("inner machanism" I guess). Despite this time I cannot give a higher score, I would like to encourage the authors to continue it and publish it in near future. Thank you author(s).

Reviewer_Us9WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how musical style is represented inside a large music generation model, and which components contribute to that representation, using circuit analysis. Targeting YuE2, the authors analyze how style is represented and formed, drawing on the analysis methodology used in the prior work Emotion Circuits. They confirm that the representation of musical style changes depending on the token position during score generation, and that the important components also differ from position to position. To address this property, they propose a method that aggregates activation differences, using as weights the KL divergence between the output distributions with and without the style specification at each position. This aggregation of a dynamic style representation is the main difference from existing methods. In their experiments, the authors identify a sparse circuit consisting of the attention heads and MLP neurons that contribute to the style representation, and verify its causal contribution through ablation and activation enhancement. Furthermore, they show that intervening on the identified components can steer the generated music toward a target style without using any text that specifies the style.

Critical strengths
  • The method in this work largely follows the framework of Emotion Circuits, so the novelty of the methodology itself is limited. On the other hand, the paper deserves credit for experimentally revealing that, in music generation models, the internal representation of style and its important components change depending on the token position, and for proposing a reasonable solution to this in the form of KL divergence weighted aggregation. Although the only target of the experiments is YuE2 and the generality of the findings leaves room for verification, the results obtained are not trivial. The paper provides new insights into the internal mechanisms of music generation models, and I consider it to make a meaningful contribution to the community.
Critical weaknesses
  • Whether this should be regarded as a critical weakness is debatable, but the fact that the analysis is limited to YuE2 alone remains a concern when considering the generality of the findings. The prior work Emotion Circuits also targeted only two LLMs, and I understand that verification across multiple models is not easy in mechanistic interpretability research. However, it is not clear whether the observation that the style representation changes with token position, or the effectiveness of the KL weighted aggregation, depends on properties specific to YuE2.
  • Regarding the comparison between CAA and the proposed method, it is a concern that the success rate and the truncation rate for each intervention strength are not sufficiently presented. The proposed method achieves a higher success rate and a lower truncation rate than CAA, but this difference may depend on a specific strength setting.
  • While the proposed method outperforms the style prompt in the automatic evaluation, the same advantage is not confirmed in the subjective evaluation by humans. Although the number of evaluators is small at eleven and the difference is not statistically significant, it remains unclear whether the improvement in the automatic evaluation translates into an improvement in the musical style perceived by humans. This point does not require additional experiments, but I believe it should be taken into account when interpreting the evaluation.
  • I have a concern about the choice of evaluation metrics. When analyzing the contribution of each generation stage, the authors use the CLAP margin, whereas in the main evaluation they report only a binary success judgment of whether the CLAP similarity to the target style is higher than that of the other styles. However, this metric cannot assess how clearly the target style is expressed relative to the other styles. For example, even when the similarities of the target style and the runner up style differ only slightly, the case is judged a success, so a high success rate does not necessarily imply a strong fit to the style. In fact, the subjective evaluation by humans does not show the proposed method outperforming the style prompt, which may be related to this limitation of the evaluation metric. I therefore hope the authors will report the mean and distribution of the CLAP margin for each method, and compare not only the success rate but also the strength of the style fit.
Questions for authors
  • Do authors think that the position dependence of the style representation and the effectiveness of the KL weighted aggregation generalize to music generation models other than YuE2? In particular, is there any reason to believe that the same property would be observed even in models without score planning?
  • For CAA and the proposed method, can you report the success rate and the truncation rate when varying the intervention strength? I would like to confirm that the advantage of the proposed method does not depend on a specific strength setting.
  • For each method in Table 1, can you report the mean and distribution of the CLAP margin? I would like to compare the fit to the target style rather than only the binary success rate.
  • While the proposed method is superior in the automatic evaluation, its advantage over the style prompt is not confirmed in the human evaluation. How do you interpret this difference in results?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Us9W (1/2): W1, W2

W1. Limited to YuE2

Thank you for your understanding. We added experiments on LeVo 2, which was trained by a different team, has a different architecture and no score planning. Applying the revised method as is, on 15 full songs from the official test set and without any style text, the circuit's style hits clearly exceed those of the paper's style prompt under all three judges (General Response, 3.1). Whether position dependence and the effectiveness of KL weighting depend on properties specific to YuE2 is addressed in Q1.

W2. The CAA comparison may depend on the strength setting

Thank you for this suggestion. We first reselected the steering layer and strength for CAA, following the setting of Table 1 with CAA replacing only the score-stage intervention. Strength is the total norm written into the residual stream, with the score-stage circuit as 1×; selection used only the first seed set (75 samples per setting):

SettingCLAPMuQ-MuLanTruncationMean duration (s)
CAA layer 4, 2× (submitted-version setting)596011161
CAA layer 9, 2×485022184
CAA layer 14, 1×5960053
CAA layer 14, 2× (selected)6171062
CAA layer 14, 4×61646149
CAA layer 20, 2×6162062

The paper's setting (layer 4, 2×) was neither the best nor free of truncations. After re-tuning (layer 14, 2×), CAA's truncations drop to 0, but on the second seed set, which took no part in selection, its hits remain below the circuit's (69 / 67 vs. 63 / 58).

With CAA fixed at layer 14, we then compared four strength levels separately in the score stage and the music semantic token stage: when one stage is controlled, the other receives no intervention; both methods have the same strength at every level; each setting has 75 samples, and truncations count as misses.

StageStrengthCircuit CLAP / MuLan / FlamingoCAA CLAP / MuLan / FlamingoTruncation (circuit / CAA)
Score0.5×18 / 19 / 2414 / 17 / 142 / 0
Score1×42 / 36 / 4618 / 17 / 162 / 0
Score2×38 / 45 / 5623 / 24 / 2613 / 2
Score4×25 / 31 / 3517 / 15 / 2215 / 7
Music semantic token0.5×35 / 26 / 3530 / 22 / 320 / 0
Music semantic token1×54 / 41 / 6351 / 34 / 470 / 0
Music semantic token2×65 / 71 / 7151 / 59 / 580 / 1
Music semantic token4×41 / 42 / 3433 / 47 / 351 / 0

Without control, each judge gives 15 hits. In the score stage, the circuit is never below CAA and is significantly better at most strengths (McNemar test); in the music semantic token stage, it is better at 1–2× and comparable at the extremes. The cost is that, in the score stage, the circuit truncates more often at 2× and above; the revised version will state this.

AuthorsOpenReview ↗

Response to Reviewer Us9W (2/2): W3, W4, Q1–Q4

W3. Human evaluation does not confirm the advantage

We agree, and the paper also states this (§4.5 and the third point of the limitations): in the preliminary human evaluation, both the circuit and the style prompt are clearly above the empty-tag baseline, and the circuit matches the style prompt. The revised version will restrict "outperforms the style prompt" to CLAP and MuQ-MuLan; for why the two evaluations disagree, see General Response, point 4.

W4. Only binary hits reported

Thank you for this suggestion; binary hits indeed cannot reflect how strongly a style is expressed. We added the style margin for each condition in Table 1 (the target-style score minus the highest score among the other four styles). Since the two judges have different scales, comparisons are made only within the same judge; intervals are from a cluster bootstrap over songs with the same or reworded lyrics.

ConditionCLAP mean (95% interval)MuQ-MuLan mean (95% interval)
Blank−0.142 (−0.163, −0.125)−0.197 (−0.219, −0.178)
Full prompt+0.069 (+0.051, +0.082)+0.129 (+0.113, +0.141)
CAA (re-tuned on the test set)+0.091 (+0.086, +0.099)+0.156 (+0.138, +0.178)
Submitted method+0.106 (+0.091, +0.117)+0.173 (+0.150, +0.192)
Revised circuit+0.133 (+0.115, +0.146)+0.214 (+0.190, +0.231)

In item-by-item pairing, the circuit's margin is higher than that of the full prompt on about 70% of items (Wilcoxon p < 10⁻⁸), and it is also higher relative to CAA. In the empirical cumulative distribution of the margin (Figure R2), the circuit's curve lies to the right of the other conditions overall. The circuit's high hit rate is therefore not built up from marginal hits.

Figure R2. Empirical cumulative distribution of the style margin for each condition in Table 1 (left: CLAP; right: MuQ-MuLan; 150 samples per condition). The further right a curve lies, the larger the margin; dots mark the medians, the solid vertical line marks a margin of 0, and the two dashed lines mark the upper bound of marginal hits and the lower bound of clear hits.

For generation quality, see our response to Reviewer 4oTW (W1); for Music Flamingo, see General Response, 3.4.

Q1–Q4

  • Q1, models without score planning: the two points need separate answers. First, the KL-weighted circuit transfers: LeVo 2 has a single generation stage, corresponding to YuE2's music semantic token stage, and the method applied as is clearly outperforms the paper's style prompt (General Response, 3.1). Second, position dependence also appears on LeVo 2: its style difference vectors are only partly consistent across position segments, and KL is concentrated in a few early positions.
  • Q2, success and truncation at each strength: see W2 (four strength levels in each stage, hits under three judges and truncations).
  • Q3, mean and distribution of the margin: see W4; under both judges the revised circuit has the highest mean margin.
  • Q4, automatic vs. human evaluation: see the joint explanation in General Response, point 4. In short: the gap lies in classic funk; CLAP may underestimate the style prompt; the reference examples came from style-prompt generations; the excerpts are short and the listeners few.

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 for the detailed response and the additional experiments.

The margin distributions (W4) and the comparison across strengths (W2) address my concerns. I also agree with limiting the claim about exceeding the style prompt to CLAP and MuQ-MuLan, given the human evaluation results.

I appreciate the LeVo 2 experiment, since it deals with my question about generality. I agree with Reviewer 4oTW’s concerns about the strength of the prompt baseline and the missing WER and quality metrics, and I look forward to your reply on those points. In addition, I have two short questions.

  • Do you have quantitative results on position dependence in LeVo 2? If the KL is concentrated in a few early positions, I wonder whether a single-position representation would already be sufficient there.
  • Is there a comparison with random components on LeVo 2, as in the YuE2 experiments?

I understand that time is limited, so results you already have would be enough.

AuthorsOpenReview ↗

Follow-up Response to Reviewer Us9W: position dependence and random components on LeVo 2

Thank you for your positive feedback. For the two points you share with Reviewer 4oTW (the strength of the LeVo 2 prompt baseline, and the WER and quality metrics), please see W4 in our Follow-up Response to Reviewer 4oTW (3/3). Below we briefly answer your two questions:

Condition (LeVo 2)CLAPMuQ-MuLan
Blank3429
Style prompt5467
Random components3230
Single-position circuit106117
KL-weighted circuit126134

Position dependence. On LeVo 2, the circuit built from the single-position representation already achieves most of the style control, but its hits remain significantly lower than those of the KL-weighted circuit, with the gap concentrated in funk and blues. Thus, although position dependence is weaker on LeVo 2, KL weighting is more comprehensive across styles.

Random components. Following the YuE2 experiments, random components in the same layers, of the same type and with matched injection strength, give the same hits as the blank condition, showing that the style control comes from the components selected by the circuit.

If these additions answer your questions, we would be grateful if you would take them into account in your final assessment. Thank you again for your review.

Reviewer_4oTWOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how genre is represented in YuE2, which first writes an ABC score, then music tokens, then audio. A stage-wise ablation of the style tag attributes about 76% of the style effect to the score stage. Style directions extracted at different score positions are nearly orthogonal, which the authors interpret as a dynamic representation. They then aggregate the activation differences between the style tag and the empty tag across positions, weighted by the per-step KL divergence, select 560 neurons and heads that write along the result, and inject the difference into them without any style text. On 150 test generations, this circuit reaches 133 CLAP and 126 Gemini hits, compared with 113 and 111 for the style prompt.

Critical strengths

A carefully controlled analysis of where style forms in a multi-stage music model. The Shapley decomposition of the stage-wise ablation (Fig. 2) and the cross-position comparison (Fig. 3) give clear answers to where and when the style tag acts in YuE2.

Critical weaknesses
  1. No content-preservation analysis. The lyrics are the only input kept fixed, yet the paper never checks whether they are still sung, e.g. by WER on transcribed vocals. The method appears to hurt lyric adherence, as apperant in their own shared examples - "Desert well"; For "pop", "Metal" the first line of the lyrics is generated wrong. In addition there are not generation quality metrics to relay the impacts on the generation quality itself (FAD, Aesthetic scores). This lack of evidence makes the validity of this work questionable.

  2. Unclear choice of judges for the main result. Table 1 relies on zero-shot CLAP and Gemini 3.6 Flash, neither validated for genre classification. Why not report accuracy under an off-the-shelf state-of-the-art genre classifier? Gemini is an odd choice; a dedicated music LLM such as Music Flamingo would be easier to justify. Both judges also filtered the data the steering directions were extracted from (App. A), so the circuit may simply reproduce what these judges reward. The only independent check, a preliminary listening test, slightly favors the prompt (picked as closest 27 vs. 14 times). As it stands, Table 1 is not very informative.

  3. Not style transfer, and unclear merit. The abstract claims the circuit "transfers one piece into the target style," but every condition writes a new score, and Table 13 shows the melody is no better preserved than with a different seed. Only the lyrics are kept, so this is genre-conditioned generation. Its benefit over simply prompting is unclear: the injected difference is derived from the style tag itself, and a simpler residual-stream injection without a circuit comes close (60 vs. 65 of 70 CLAP hits, Table 9).

  4. Limited novelty beyond the analysis. The component selection, ablation/enhancement validation and budget allocation follow Emotion Circuits (Wang et al., 2025a), and the steering vector follows CAA (Rimsky et al., 2024). The KL weighting is the only new ingredient. The "circuit" is a ranked set of components with no account of how they interact. The analysis itself is tied to YuE2's two-stage design and tested on no other model, so it is unclear what transfers to other music generation models.

Questions for authors

Questions

  1. Is the goal style transfer of an existing piece or genre-conditioned generation? If the latter, what is the advantage over prompting?
  2. How do the Table 1 conditions compare under an off-the-shelf genre classifier not used for data filtering?
  3. What is the lyric WER of each condition? could you provide further generation quality analysis? (FAD, Aesthetics scores)

Suggestions

  1. State the task explicitly in §3.1 and align the abstract with it.
10 comments
AuthorsOpenReview ↗

Response to Reviewer 4oTW (1/3): W1, W2

W1. No content-preservation analysis

Thank you for pointing out this gap. To separate style from content, we added score-preserving style transfer: the score is kept unchanged and the circuit steers only the music semantic token stage (C_c), and most of the melody is preserved (see General Response, point 2). Demo: Score-Preserving Style Transfer.

Following your suggestion, we also added lyric and generation-quality metrics for the conditions in Table 1 and the two score-preserving conditions (150 samples per condition): following the YuE2 technical report, we transcribe the full mix with Qwen3-ASR-1.7B (Shi et al., 2026) and compute WER; audio quality uses Audiobox Aesthetics (Tjandra et al., 2025) PQ and CE (out of 10); FAD (Kilgour et al., 2019) uses VGGish (Hershey et al., 2017) embeddings with real recordings of the corresponding GTZAN genre as the reference, lower being closer (GTZAN has no funk, so FAD uses the other four styles).

ConditionHits (CLAP / MuLan)WERPQCEFAD
Blank30 / 300.098.337.642.39
Full prompt113 / 1210.078.237.621.53
CAA (re-tuned on the test set)124 / 1290.157.957.361.31
Submitted method133 / 1270.178.157.521.42
Revised circuit (C_s C_c)142 / 1390.447.857.261.55
Fixed score, tag only (P_c)43 / 410.118.347.662.01
Fixed score, circuit (C_c)109 / 990.177.987.341.26

Overall, the circuit sings slightly more words wrong or omits slightly more words than the corresponding prompt, and its audio quality is slightly lower, mainly due to metal rock. Automatic transcription itself also introduces errors: even the blank renders, generated without any intervention, have a WER of 0.09, and the transcriber mishears some words that are sung correctly (see "Desert well" below and the per-sample check). FAD, which reflects both style and audio quality, still shows the circuit's advantage: with a fixed score, C_c is clearly closer than P_c to real recordings; when generating from lyrics, C_s C_c matches the full prompt. Excluding metal rock, the audio-quality gap shrinks to within 0.2 points, the circuit still has slightly more lyric errors, and the FAD of C_s C_c is slightly worse than that of the full prompt (full table in the revised appendix).

On "Desert well": we listened to each sample. The metal rock version from the submitted method does omit "the bucket" in the first line, while the full prompt and the empty-tag baseline with the same lyrics and seed sing it correctly; the pop version sings it correctly, but Qwen3-ASR transcribes "the bucket" as "the boat". Per-sample audio and transcriptions: Lyric Transcription Check.

W2. Unclear judge choice for the main results

Thank you for this comment. We added two judges that were not involved in data filtering: MuQ-MuLan (Zhu et al., 2025; the text–audio contrastive model used in the YuE2 technical report) and the music LLM you suggested, Music Flamingo (Ghosh et al., 2025). Music Flamingo picks one of the five style labels; we rotate the order of the options five times and average, to remove its preference for particular option positions.

JudgeUsed in filteringRevised circuitSubmitted methodFull promptBlank
CLAPYes14213311330
MuQ-MuLanNo13912712130
Music FlamingoNo14413813830

With judges not involved in filtering, the conclusion agrees with the paper's claim: the circuit matches and even exceeds the style prompt (it exceeds the prompt under MuQ-MuLan and matches it under Music Flamingo). In the score-preserving setting, which the prompt cannot control, and on LeVo 2, all three judges show the circuit clearly stronger (General Response, points 2 and 3.1). In addition, the revised circuit is fitted only on histories written with an empty tag and does not use the filtered style audio. For the human evaluation, see General Response, point 4.

AuthorsOpenReview ↗

Response to Reviewer 4oTW (2/3): W3, W4

W3. Not style transfer, and its value is unclear

Task Problem Statement: When generating from lyrics, the experiments in the submitted version indeed keep only the lyrics and not the score; a more accurate description is "fixed lyrics, steering generation toward a target style". However, steering generation toward a target style directly inside the model, without preserving the content of a given piece, has also been studied as style control in recent work: for example, Composer Vector (Jiang et al., 2025) adds a composer style vector to the residual stream of a symbolic music model to generate new pieces in that composer's style. The submitted version starts from a similar perspective; moreover, genres already differ at the level of the score, so rewriting the score is itself a composition-level change of style (see General Response, point 2). Beyond, our method can also perform style transfer that preserves the score (next paragraph). The revision will clearly distinguish these two settings in the abstract and §3.1 (Suggestion 1).

Score-preserving style transfer: we extended KL weighting to the music semantic token stage, so the circuit can steer this stage alone: the score is fixed and only the music semantic tokens change; this is exactly the style transfer you describe. The circuit then clearly changes the style, whereas the tag alone at the same stage barely changes it (all three judges agree), and most of the melody is preserved, with the loss concentrated in metal rock. Results and melody analysis are in General Response, point 2. Demo: Score-Preserving Style Transfer.

Comparison with the prompt: The injected difference does come from the style tag; changing only the tag and comparing the activations of the two forward passes is the standard way of constructing counterfactuals in interpretability analysis. This difference reveals how the model represents style internally and stably localizes the circuit. In application, once the score is fixed the prompt can barely change the style, whereas the circuit controls style directly at the music semantic token stage, enabling melody-preserving style transfer.

On residual-stream steering (Table 9, 60 vs. 65): Table 9 has only 70 samples. We repeated this comparison on new test lyrics (both methods steering only the score stage): under all three judges the circuit has more hits and a significantly higher margin, with no difference in lyrics or audio quality; see our response to Reviewer fUFC (W2).

W4. Limited novelty beyond the analysis

Positioning of the contributions (General Response, point 1): component selection and validation do follow Emotion Circuits; the novelty lies not in this pipeline but in the stage and position analysis, the KL-weighted cross-position aggregation derived from it, and the circuit analysis of the aggregated representation.

(1) Other models: We applied the revised method to LeVo 2 as is (a different team, a different architecture, a single generation stage). On 15 full songs from the official test set, without any style text, the circuit's style hits clearly exceed those of the paper's style prompt under all three judges (General Response, 3.1). Circuits fitted on real recordings are equally effective (General Response, 3.5).

AuthorsOpenReview ↗

Response to Reviewer 4oTW (3/3): W4 (continued), Q1–Q3, Suggestion 1

W4. Limited novelty beyond the analysis (continued)

(2) How components interact: As you point out, following Emotion Circuits, "circuit" means the sparse set of components that contributes most to the style representation, and the submitted version did not analyze interactions among components. We added two analyses. First, ablations of the number and type of components, with each retained component keeping its strength:

ComponentsMLP neuronsAttention headsCLAPMuQ-MuLan
Full circuit392168140137
Half19684133128
Double784336139140
Attention heads only0168132132
MLP neurons only39203434
Blank——3030

The attention heads account for most of the style effect: heads alone reach over 90% of the full circuit's hits, while neurons alone are close to blank; halving or doubling the components changes the hits little.

Second, we analyzed the attention heads selected by the circuit (Figure R1). Ablating attention to the tag in only these 168 heads leaves about 40% of the tag's effect on the output distribution, while masking the same number of random heads has almost no effect, so the tag reaches later positions mainly through the circuit heads. The circuit heads also share a division of labor: some read the tag directly, while more of them attend to style-related score content such as the key signature.

Figure R1. (a) Fraction of the tag's effect on the output distribution (KL) that remains after ablating attention to the tag in the circuit heads or in the same number of random heads; (b) proportions of several score-stage roles among circuit heads and random heads; (c) attention of L21H1, a head specific to the blues circuit, to the key signature and note durations under the empty tag and the blues tag.

(3) The circuit integrates information across positions: The KL-weighted representation is aggregated and is never directly produced by the model at any single position, so the components that consistently write it should integrate information from different positions. Component selection shows this directly. We repeat the circuit's component selection, with the same rule and the same number of components, such as the key field of the header or the vocal pitches in the early part of the body; this gives 22 component sets (4 header fields, plus 6 body token types, each in the early, middle and late part). Each attention head of the circuit appears in about 17 of these 22 sets on average (about 10 for random heads with the same layer distribution), whereas heads selected from the first position of the score alone appear in only about 9, close to random.

Q1–Q3 and Suggestion 1

  • Q1, style transfer or genre-conditioned generation: the main experiment of the submitted version steers generation toward a target style without preserving the score, which matches the perspective of recent work on style steering inside the model (e.g., Composer Vector); on top of this, steering only the music semantic token stage gives score-preserving style transfer, where the tag alone has little effect. See W3 and General Response, point 2. Demo: Score-Preserving Style Transfer.
  • Q2, off-the-shelf classifiers: under MuQ-MuLan the revised circuit exceeds the full prompt (139 vs. 121), and under Music Flamingo it matches the style prompt (144 vs. 138); with a fixed score or on LeVo 2, all three judges show the circuit clearly stronger. See W2.
  • Q3, lyric WER and generation quality: see W1. The circuit has slightly more lyric errors and slightly lower audio quality, mainly in metal rock; on FAD it matches the full prompt when generating from lyrics and is clearly better than the tag alone with a fixed score.
  • Suggestion 1: We adopt it. The abstract and §3.1 will describe the task as "style-conditioned generation with fixed lyrics" and report score-preserving style transfer separately; "matching and even exceeding the style prompt" in the abstract will be qualified as "under CLAP and MuQ-MuLan", stating that under Music Flamingo and in the human evaluation the circuit matches the style prompt.

The references cited above are listed at the end of the General Response.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response and Summary for the AC (1/4): Contribution and discussion outcome

Dear AC and all reviewers ,

Thank you for handling our paper. And thanks for all reviewers for their detailed comments! Below we summarize the contribution, what the rebuttal added, and where the discussion with each reviewer stands; points 2–4 of the General Response give the details.

1. Main contribution

To our knowledge, this is the first circuit-level interpretability study of musical style in a large music generation model. We highlight two key steps. First, stage and position analyses show that the style representation is dynamic: in YuE2 it forms mainly in the score stage, and no single position represents it. Second, KL weighting, a simple and effective scheme, merges the style tag's effect across positions into one direction per style. The sparse circuit of attention heads and MLP neurons that writes this direction steers style without any style text and matches the style prompt.

2. What the rebuttal added

  • Score-preserving style transfer: with the score fixed, steering only the music semantic token stage clearly changes the style, which the tag alone barely does, and preserves most of the melody.
  • Another model: the method transfers unchanged to LeVo 2 and clearly exceeds the style prompt without any style text.
  • Independent judges: under MuQ-MuLan, Music Flamingo and the supervised classifier MAEST, none used in data filtering, the circuit matches or slightly exceeds the style prompt.
  • Lyrics and quality: WER and Audiobox show a cost, mainly in metal rock; FAD against 5,548 FMA-large recordings shows no difference from the prompt.
  • Real recordings: circuits fitted on GTZAN recordings also steer three new genres. We added this ourselves as further evidence of generality; no reviewer requested it.

3. Discussion outcome

fUFC agreed that the added ablations support KL weighting and clarify what component localization adds, and that the stage and position analyses add value beyond it. The reviewer keeps weak accept, viewing the methodological contribution as modest.

Us9W found that the margin distributions and the comparison across strengths address the concerns, and agreed with limiting "exceeding the style prompt" to CLAP and MuQ-MuLan. We answered the two follow-up questions on LeVo 2; unfortunately, the reviewer has not replied further.

dHzM. Our response added independent judges and fixed-score results; And finally the reviewer keeps the score.

4oTW noted that the claims are now stated more accurately, and welcomed the component analysis and the LeVo 2 experiment. The reviewer keeps the score, as the drop in lyric accuracy and audio quality and the small gains over prompting remain. We accept these points and will state them as limitations. We respectfully disagree only on whether the fixed-score setting is a new method: it uses the same KL weighting, component selection and injection, applied at the music semantic token stage.

AuthorsOpenReview ↗

General Response (2/4): Score-preserving style transfer

4. Score-preserving style transfer: the same method at another stage

Why this is not a new method. The submission already analyzes both token stages of YuE2: the style tag acts mainly through the score (Shapley share 76%), and the music semantic token stage carries the rest. The KL weighting, the component selection and the injection do not depend on the stage and are used here unchanged. The only difference from the main experiment is that the score is kept fixed and the circuit is applied at the music semantic token stage.

What it adds. This setting tests a prediction of the stage analysis: once the score is fixed, the style prompt should lose most of its effect, while a circuit that writes the style direction directly should not. It also separates style from content. Following the hierarchical definition of Dai et al. (2018), steering the score keeps only the lyrics, which, as 4oTW pointed out, is genre-conditioned generation; steering only the music semantic tokens keeps the score and changes the arrangement, singing and timbre, which is melody-preserving style transfer.

Below, P denotes giving the style tag at a stage and C denotes steering with the circuit at that stage; subscript s is the score stage and c is the music semantic token stage.

Intervened stage(s)Source of the scoreCLAPMuQ-MuLanMusic FlamingoTag only at the same stage (CLAP / MuLan / Flamingo)
Score only C_s (submitted-version setting)Generated with the circuit77699587 / 86 / 106
Music semantic tokens only C_cFixed to the empty-tag score1189513351 / 44 / 53
Both stages C_s C_cGenerated with the circuit142139144Full prompt: 113 / 121 / 138
Empty-tag baseline—303030—

As the stage analysis predicts, with the score fixed the tag alone barely changes the style (51 / 44 / 53), whereas the circuit clearly does (118 / 95 / 133). Demo: Score-Preserving Style Transfer

Melody preservation. Following the score–audio protocol of the YuE2 technical report (§5.1), we transcribe the generated audio with SheetSage2 (Jiang et al., 2026) and compare its vocal melody with the fixed score (melody similarity: one minus the normalized edit distance between the pitch sequences; rhythm: F1 of consecutive onset intervals). As a control, the same audio is also compared with another score written for the same lyrics.

Condition (fixed score)Style hits (CLAP / MuLan)Melody similarityRhythm F1
No control30 / 300.960.92
Tag only (P_c)43 / 410.950.92
Circuit (C_c)109 / 990.850.82
KL-weighted residual steering (layer 14, 2×)129 / 1380.410.42
Equal-weight CAA (layer 14, 2×)105 / 1210.420.44
Random components31 / 310.960.93
Control: no-control audio vs. another score—0.090.49

The circuit's audio stays far closer to its own score than to another score, with the loss concentrated in metal rock. The tag alone keeps the melody but not the style, single-layer residual steering gains style but loses most of the melody, and random components change neither; only the circuit achieves both.

AuthorsOpenReview ↗

General Response (3/4): Additional experiments

5. Summary of additional experiments

5.1 LeVo 2 (4oTW, Us9W). We applied the music-semantic-token-stage method unchanged to LeVo 2 (Lei et al., 2025), a model from a different team with a single generation stage and no score planning, and tested it on the 15 full songs of the official test set (150 samples per condition).

ConditionCLAPMuQ-MuLanMusic Flamingo
Blank342927
Style prompt (the paper's style tags)546787
Official prompt format (genre + instruments)607595
Circuit126134143

Without any style text, the circuit clearly exceeds both prompts, but it also raises WER (0.28 vs. 0.11). Samples: LeVo 2 Full-Song Samples

5.2 Style strength (Us9W). Under CLAP and MuQ-MuLan, the circuit's style margin is significantly higher than that of the full prompt (response to Us9W, W4).

5.3 Lyrics and audio quality (4oTW). WER rises from 0.07 (full prompt) to 0.17 (submitted method) and 0.44 (revised circuit), and Audiobox quality drops, mainly in metal rock; FAD against 5,548 FMA-large recordings shows no difference from the prompt. The revision will report these per style and state the trade-off (responses to 4oTW, W1).

5.4 Independent judges (4oTW, dHzM). Under MuQ-MuLan, Music Flamingo and MAEST (Alonso-Jiménez et al., 2023), none used in data filtering, the circuit matches or slightly exceeds the prompt when generating from the lyrics, and clearly exceeds it with a fixed score and on LeVo 2 (responses to 4oTW, W2).

5.5 Real recordings (added by us; not requested by any reviewer). A circuit refitted with the same KL-weighting formula on 100 real GTZAN recordings steers reggae, jazz and disco with the score fixed: 71 / 62 CLAP / MuQ-MuLan hits out of 90, against 6 / 6 for random components and 84 / 71 for the circuit fitted on generated data. Samples: Circuits Fit on Real Music

AuthorsOpenReview ↗

General Response (4/4): Human evaluation; references

6. Automatic vs. human evaluation (4oTW, Us9W, dHzM)

In the preliminary human evaluation (11 listeners), both the circuit and the style prompt are clearly above the empty-tag baseline, and the circuit matches the style prompt; the difference is not significant. The gap lies in a single style, classic funk, and the style examples that listeners heard beforehand were taken from style-prompt generations (Appendix D.3), which can only favor the prompt. The revision will therefore limit "exceeding the style prompt" to CLAP and MuQ-MuLan, and state in §4.5 and the limitations that the circuit matches the prompt in the human evaluation.

References

  • Alonso-Jiménez, P., Serra, X., & Bogdanov, D. (2023). Efficient Supervised Training of Audio Transformers for Music Representation Learning. ISMIR.
  • Dai, S., Zhang, Z., & Xia, G. G. (2018). Music Style Transfer: A Position Paper. International Workshop on Musical Metacreation (MUME).
  • Defferrard, M., Benzi, K., Vandergheynst, P., & Bresson, X. (2017). FMA: A Dataset for Music Analysis. ISMIR.
  • Ghosh, S., Goel, A., Koroshinadze, L., et al. (2025). Music Flamingo: Scaling Music Understanding in Audio Language Models. arXiv:2511.10289.
  • Hershey, S., Chaudhuri, S., Ellis, D. P. W., et al. (2017). CNN Architectures for Large-Scale Audio Classification. ICASSP.
  • Jiang, J., Yuan, R., Pan, J., Xue, W., Guo, Y., Xia, G., & LeCun, Y. (2026). SheetSage2: Coherent Lead-Sheet Transcription with Synthetic Supervision. arXiv:2610.05336.
  • Jiang, X., Xu, X., & McAuley, J. (2025). Composer Vector: Style-steering Symbolic Music Generation in a Latent Space. NeurIPS 2025 Workshop on AI for Music.
  • Kilgour, K., Zuluaga, M., Roblek, D., & Sharifi, M. (2019). Fréchet Audio Distance: A Reference-Free Metric for Evaluating Music Enhancement Algorithms. Interspeech.
  • Lei, S., et al. (2025). LeVo: High-Quality Song Generation with Multi-Preference Alignment. NeurIPS.
  • Shi, X., Wang, X., Guo, Z., Wang, Y., Zhang, P., Zhang, X., Guo, Z., Hao, H., Xi, Y., Yang, B., Xu, J., Zhou, J., & Lin, J. (2026). Qwen3-ASR Technical Report. arXiv:2601.21337.
  • Tjandra, A., Wu, Y.-C., Guo, B., et al. (2025). Meta Audiobox Aesthetics: Unified Automatic Quality Assessment for Speech, Music, and Sound. arXiv:2502.05139.
  • Zhu, H., et al. (2025). MuQ: Self-Supervised Music Representation Learning with Mel Residual Vector Quantization. arXiv:2501.01108.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.