acceptodds
Under review as a conference paper at ICLR 2027

Plan, Generate, Replan: Hierarchical Symbolic Music Generation with Boundary-Aware History Correction

Abstract

Music creation is an iterative process: composers repeatedly revise what has been written and decide how the music should develop next. Supporting this process requires modeling how music evolves under user instructions. We introduce the Music World Model (MWM), a hierarchical latent world model for controllable long-form symbolic music generation. MWM represents musical content as latent states and requested changes as explicit actions. A phrase-level planner predicts structural context, while a segment-level executor recursively generates musical states under this context and the requested action. Users can update actions at phrase boundaries, and the planner refreshes its predictions from generated history to account for how the music has evolved. However, mismatches between new plans and preceding context can enter this history and influence subsequent plans, causing structural drift. We therefore introduce Boundary-Aware History Correction (BHC), an inference-time mechanism that combines context anchoring with persistent correction memory to reconcile planned slow trajectories with preceding context. BHC carries correction information across phrase boundaries while keeping learned weights and external actions fixed, without future reference music. To support learning these action-conditioned dynamics, we construct Slakh-Action from Slakh2100 Redux, aligning multi-track symbolic music with segment-level states, phrase-level summaries, and structural action annotations. An anonymous interactive demo is available at https://demo-czv.pages.dev/

Primary area:
generative models
Keywords:
Latent dynamics, Hierarchical planning, Controllable music, Long-horizon generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iJFkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Music World Model, a hierarchical framework for controllable long-form symbolic music generation. MWM combines a phrase-level planner with a segment-level executor to model musical evolution under structural actions, allowing the model to replan based on generated history. To mitigate structural drift during recursive generation, the authors introduce Boundary-Aware History Correction (BHC), an inference-time mechanism that combines context anchoring with persistent correction memory. The authors also construct Slakh-Action to support action-conditioned training and evaluation, and conduct experiments to demonstrate the effectiveness of the proposed framework.

Critical strengths
  • The paper addresses the important challenge of maintaining structural consistency and controllability in long-form symbolic music generation. The proposed plan–generate–replan framework provides an interesting perspective on interactive music generation. -The paper presents a relatively complete technical framework, integrating hierarchical planning, action-conditioned generation, and inference-time history correction. The overall design is coherent, and the individual components are motivated by specific challenges in long-horizon generation.
  • The authors conduct a broad range of experiments, including comparisons with existing methods, subjective evaluations, counterfactual action tests, long-horizon generation, and ablation studies. These experiments provide useful insights into the behavior of the proposed framework.
Critical weaknesses
  • While the paper presents the proposed approach as a Music World Model, it is not entirely clear how the underlying task differs from conventional conditional symbolic music generation. The framework predicts future musical content based on historical states and structural controls, which seems closely related to existing conditional generation approaches. Furthermore, the training actions are derived from structural differences between consecutive phrases in existing music, rather than from actual user interventions. This raises questions about whether these inferred actions accurately reflect real user intentions and whether the learned dynamics can effectively respond to user-specified interventions in practical scenarios. how the proposed framework demonstrates the predictive and generalization capabilities expected of world models, beyond conditional continuation with predefined structural attributes.

  • In Table 9, the full MWM model performs worse than several simplified variants, including the No-plan and Executor-only configurations, across multiple evaluation metrics. But the authors didnt discuss these results.

  • The relationship between the action and the control descriptors could be explained more clearly. In particular, how would the model handle situations where the action and control descriptors specify conflicting changes?

  • how the proposed method differs from existing approaches to long-form symbolic music generation, particularly Anchored Cyclic Generation (ACG) and its hierarchical extension.

  • While the paper presents a relatively comprehensive technical framework, some parts of the presentation could be simplified to improve readability. The large number of concepts, representations, symbols, and technical terms makes it challenging to follow the relationships between different components and grasp the main methodological ideas. Additionally, some terminology and notation appear inconsistent between the main text and the appendices.

Questions for authors

refer to the weakness

3 comments

1. World-model distinction.

MWM explicitly models musical state–action transitions and replans from generated results, examining how the same starting point develops under different interventions. Phrase differences supervise musical changes during training as proxies for interventions, not direct observations of user intent, while our agent translates user intentions into tempo, density, and role-activity controls through the same action interface. Instruction comparisons and multi-step prediction experiments on held-out songs evaluate intervention responses and predictive ability. Generalization here means applying learned controls to unseen musical contexts. The anonymous link provides complete code and configurations for inspecting intent interpretation and generation.

2. Planner (Table 9).

We acknowledge that Executor-only and No-plan achieve better results on several metrics. This is mainly because the planner introduces additional constraints on how the music develops, while removing it gives the executor more freedom to generate music that matches local statistical patterns, leading to better scores on some metrics. However, our goal in introducing the planner is not simply to improve these scores, but to organize how music develops across multiple phrases. Without a planner, the model can maintain good scores by repeating similar chords or rhythmic patterns, but this can also make longer pieces sound repetitive. In contrast, the planner helps organize changes across phrases and gives the music a clearer overall structure. Therefore, its main benefit lies in long-term musical development rather than local metrics alone. We provide audio demos to illustrate these differences, while acknowledging that current objective metrics do not directly establish this advantage. BHC is disabled in Table 9; Full(BHC) results are shown below, and we will compare the other variants with BHC enabled. We will add cross-phrase structural analyses and listening evaluations to better demonstrate the planner's value.

WindowMethodRBDE↓C-JSD↓SC↑R-JSD↓GC↑RSE↓TSSE↓
First 8Full(BHC)0.4980.3170.9850.1370.8680.1530.089
Full0.5670.3230.9850.1390.8650.1570.089
Centered 8Full(BHC)1.1040.3320.9830.2110.8680.2120.082
Full1.2040.3410.9810.2190.8610.2170.083

3. Actions and descriptors.

The user expresses an intention, which the agent translates into a structured action. The system then combines this action with the current musical state to form control descriptors for generation. For example, “make the next phrase denser” becomes a density-increase action, which is interpreted relative to the current density to establish the corresponding target. The action expresses how to change, while the descriptors express the desired state. They belong to the same control pipeline rather than two independently supplied user instructions, so conflicting action–descriptor specifications do not arise in the intended workflow.

4. ACG and Hi-ACG.

ACG supports long-sequence generation through anchor feedback, while Hi-ACG adds whole-piece sketches and local refinement. MWM centers on interaction and explicit actions: users can adjust density, tempo, or role activity at phrase boundaries, and the model predicts subsequent music from the current state and action, then replans from generated results. Users can therefore listen and revise instructions as generation unfolds, reflecting the action-conditioned prediction and feedback capabilities of a world model.

5. Presentation.

We will further simplify the technical descriptions, reduce unnecessary terminology and notation, and ensure consistency between the main text and appendices while correcting related errors. We will also add a clear and concise overview diagram to show the role of each component and how they connect, making it easier for readers to understand the main idea and overall workflow of our method. The overall framework diagram is provided in the first section of the supplementary material.

thanks for your response. i will keep my score

Thank you for your response. We respect your decision to maintain your original score. However, we would sincerely appreciate it if you could provide more specific reasons or further justification for your assessment, particularly regarding which concerns remain unresolved despite our detailed responses. Without concrete feedback, it is difficult for us to understand why our manuscript does not meet your expectations or how we can meaningfully improve our work. We believe that constructive and specific feedback is essential to the peer-review process, and we would be grateful if you could elaborate on the key issues underlying your decision.

Reviewer_DWN9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes the Music World Model (MWM), which is capable of predicting latent representations of symbolic music from a pretrained autoencoder. The prediction is further controlled by structural actions and musical attributes through a high-level planner that predicts future musical representations and controls. The authors also introduce a Boundary-Aware history Correction method to improve long-term consistency. Experiments show competitive generation quality compared to existing systems.

Critical strengths
  1. The paper explores an interesting problem of long-term controllable generation.
Critical weaknesses
  1. Although the authors motivated the MWM idea, it is still questionable whether the planner performs meaningful high-level music planning. A MWM should ideally plan toward more abstract goals than predicting low-level musical attributes (even if they are aggregated). The proposed approach seems to be a more careful version of predicting in the latent space (E.g.,https://arxiv.org/pdf/2310.16334). The paper could clarify this better.
  2. The paper involves several experiments showing the music quality, but I find them insufficient to support the motivation. The demo page results are not very convincing (also unclear; see weakness 3); they also make one understand that the reported advantages in the paper tables look like tied performance rather than clear improvements. It is also important to develop more carefully designed evaluations, e.g., accuracy of the action-conditioned dynamics, overall plan quality. Most importantly, Appendix E.1 does not demonstrate an advantage of using a planner.
  3. Presentation can be clearer. The methods involve many detailed notations without a system-level overview, introduction, or diagram. Figure 1 can be more informative than drafting a vague workflow of different approaches. The current discussion of the method is very hard to understand. The demo page instructions are very difficult to understand. I spent considerable time trying to understand them but still could not figure out what to expect from each demo. I cannot control the audio seek bar (because it is so small) and cannot see the generated MIDI. In most cases, I don't understand what is given and what is generated.

Overall, I find the technical contributions to be around Weak Reject. However, clarity issues (point 3) further lower my score to a clear rejection.

Questions for authors

I would appreciate it if the authors could address the points in the weakness section. I am most interested in understanding the accuracy of the world model itself (the action-conditioned transition) through more analysis and audio or MIDI examples. This can also help explain the low success rate in Table 3.

3 comments

1. Planning scope vs latent prediction.

We clarify that MWM's planner goes beyond predicting isolated attributes such as density. Conditioned on musical context and user instructions, it predicts future musical representations at multiple temporal scales: the next eight-bar phrase, its four two-bar segments, and pooled representations over a four-phrase horizon. The executor then converts these plans into notes. After each phrase, the planner updates subsequent plans using the actual generated music rather than following a fixed trajectory. We agree that Zhao et al. also use latent representations for long-term structure, but their focus is multi-track accompaniment arrangement of given musical material guided by style priors. In contrast, MWM targets interactive continuation, where changing user instructions require the system to repeatedly plan, generate, and replan. Thus, the distinction is not latent prediction alone, but its use in an action-conditioned, closed-loop planning–execution process. We will clarify this distinction and acknowledge that, although the four-phrase horizon extends planning beyond a single phrase, more abstract goals such as whole-piece form and emotion remain future work.

2. Planning evaluations and quality.

We agree that conventional quality metrics, the current demos, and Appendix E.1 do not directly establish the planner's benefit. We therefore introduce evaluations of action-conditioned dynamics and planning quality. Compared with No-plan, MWM increases directional accuracy from 79.16% to 83.46% and reduces 16-step state prediction error from 0.732 to 0.638. BHC further lowers this error to 0.615. Compared with Open-loop, MWM+BHC reduces planning and execution errors by 23.2% and 24.0%, respectively. These results more directly support improved long-horizon dynamics modeling and planning consistency, rather than uniformly better local quality metrics. We have added the new metrics, their definitions, and results to the supplementary material.

MethodTarget↓Dir.%↑State1↓State16↓Plan↓Exec.↓SC↑GC↑RSE↓
Executor-only0.76477.42N/AN/AN/AN/A0.9850.8630.153
No-plan0.74879.160.2630.732N/AN/A0.9850.8650.156
Open-loop0.73880.830.2510.7080.6470.6120.9850.8650.156
MWM0.72083.460.2410.6380.5230.5010.9850.8650.157
MWM+BHC0.71981.680.2430.6150.4970.4650.9850.8680.153

Regarding music quality and planning,

we acknowledge that Executor-only and No-plan achieve comparable or better scores on some metrics. Without the planner's constraints, the executor has more freedom to match local statistical patterns. Our planner instead aims to organize musical development across phrases while maintaining comparable generation quality. The new evaluations show improvements in prediction and planning, although they do not alone establish perceptually superior long-form structure. BHC was disabled in Table 9. The comparisons below show that Full(BHC) improves several quality metrics over Full in both windows.

WindowMethodRBDE↓C-JSD↓SC↑R-JSD↓GC↑RSE↓TSSE↓
First 8Full(BHC)0.4980.3170.9850.1370.8680.1530.089
Full0.5670.3230.9850.1390.8650.1570.089
Centered 8Full(BHC)1.1040.3320.9830.2110.8680.2120.082
Full1.2040.3410.9810.2190.8610.2170.083

3. Presentation and demos.

We agree that Fig. 1 emphasizes distinctions among approaches without clearly explaining MWM's end-to-end pipeline. An overall framework diagram is now provided in the first section of the supplementary material, showing each component's role and connections. We will simplify the method presentation and revise the demo with clearer instructions, explicit labels distinguishing given from generated music, more usable audio seek controls, and visualization of generated MIDI. We will also state what each demo is intended to demonstrate.

Thank you for the rebuttal. Demonstrating the effectiveness of the learned dynamics and how planning contributes to generation quality requires a more careful organization and presentation of the paper, which I believe would require a substantial revision. I am also not convinced that the writing and demo page can be sufficiently improved within the limited time after the rebuttal. Regarding the new experiments, I am still unsure how the metrics are defined and how the evaluations are designed, and I could not find the updated supplementary material. I will maintain my score.

Thank you for your continued feedback. We understand your concerns regarding the clarity of the presentation and the effectiveness of the learned dynamics and planning. However, we would like to clarify several important points that may not have been sufficiently communicated in our previous response.

First, regarding the revised framework and additional experiments: We have already provided updated supplementary material containing substantial revisions and new experimental results. Specifically, the revised framework figure is placed at the very beginning of the supplementary material, providing a clearer illustration of how learned dynamics and planning are integrated into the generation process. Furthermore, the newly added experiments, detailed metric definitions, and evaluation protocols are presented in the later sections of the supplementary material. These additions are intended to directly address your concerns about the effectiveness of our method and the clarity of the evaluation. We apologize that these updates were not sufficiently highlighted in our previous response, which may have made them difficult to locate.

Second, regarding the feasibility and extent of the revisions: We would like to emphasize that our team has dedicated at least a full week of intensive, coordinated effort to systematically addressing all the concerns raised. Through a clear division of responsibilities and parallel efforts, we have made every possible effort within the limited timeframe to implement substantive improvements, rather than merely promising future revisions. While we fully acknowledge the importance of clear presentation, we respectfully believe that the adequacy of our revisions should be assessed based on the concrete changes and evidence we have provided, rather than solely on an assumption that meaningful improvements cannot be achieved within the rebuttal period. We sincerely hope that our team's substantial efforts and the resulting improvements will receive fair consideration.

We fully respect your assessment and appreciate the time you have devoted to reviewing our work. Nevertheless, we believe that the additional experimental evidence and the substantial revisions already completed deserve careful consideration. We sincerely hope you could revisit the updated supplementary material and reconsider your assessment based on the actual improvements and evidence we have provided.

Reviewer_u2wROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes MWM to generate MIDI music under user instructions. It plans music phrase by phrase and uses BHC to reduce errors during generation. Experiments show some improvements, but the benefits of the planner and BHC remain unclear.

Critical strengths
  1. MWM allows users to change instructions during music generation and shows clear initial responses to note-density commands.
  2. BHC helps reduce errors during longer generation without retraining the model or using future reference music.
  3. The experiments cover six baselines, control tests, and ablations. MWM achieves the lowest role-density error in the main comparison.
Critical weaknesses
  1. The benefit of the phrase planner is unclear. Executor-only and No-plan perform better than the full model on several key metrics. Table9
  2. Density control is weak and becomes less reliable over time. The model responds only slightly to commands, and later responses can go in the opposite direction.
  3. Only test on Slakh dataset, which can not provide solid evidence for whether this method can be generalized to real music or music from other sources.
Questions for authors
  1. Can you further explain why do Executor-only and No-plan perform better on several metrics. When does adding the phrase planner help, is there some evidence can support its usefulness.
  2. Can you explain why the density control is weak and less reliable over time.
  3. Does the method retain its generation quality and control performance on other music datasets or styles? Can you show some results on other data, like musdb or etc.
1 comment

1. Planner (Table 9).

We acknowledge that Executor-only and No-plan achieve better results on several metrics. This is mainly because the planner introduces additional constraints on how the music develops, while removing it gives the executor more freedom to generate music that matches local statistical patterns, leading to better scores on some metrics. However, our goal in introducing the planner is not simply to improve these scores, but to organize how music develops across multiple phrases. Without a planner, the model can maintain good scores by repeating similar chords or rhythmic patterns, but this can also make longer pieces sound repetitive. In contrast, the planner helps organize changes across phrases and gives the music a clearer overall structure. Therefore, its main benefit lies in long-term musical development rather than local metrics alone. Our audio demos illustrate these differences, while we acknowledge that current objective metrics do not directly establish this advantage. BHC is disabled in Table 9; additional Full(BHC) comparisons are shown below. We will add cross-phrase structural analysis and listening evaluations to better demonstrate the planner's value.

WindowMethodRBDE↓C-JSD↓SC↑R-JSD↓GC↑RSE↓TSSE↓
First 8Full(BHC)0.4980.3170.9850.1370.8680.1530.089
Full0.5670.3230.9850.1390.8650.1570.089
Centered 8Full(BHC)1.1040.3320.9830.2110.8680.2120.082
Full1.2040.3410.9810.2190.8610.2170.083

2. Density control.

We carefully considered the reviewer's comments and re-examined our evaluation criteria. Since musical notes often occur in groups, such as chords or drum patterns, even a small musical adjustment may involve changes to multiple notes. Therefore, the original tolerance of 0.25σ (±5.71 notes per bar) was overly strict. After increasing the tolerance to 0.5σ (±11.42 notes per bar), the target response rates increased from 3.47% to 87.57% for MWM and from 3.04% to 81.82% for MWM+BHC, compared with 0.72% to 2.15% for Mask. Similarly, the reversal response rates increased from 60.03% to 97.67% for MWM and from 57.12% to 97.34% for MWM+BHC, compared with 11.29% to 14.38% for Mask. We will update these results in the paper while retaining our analysis of response sensitivity. Regarding weaker or opposite responses at later stages, the model tends to follow previously established musical patterns to maintain melodic, rhythmic, and structural coherence, reducing the influence of subsequent control instructions. Meanwhile, natural variations in musical density may counteract the requested adjustments, sometimes resulting in opposite responses. Improving long-term control stability while preserving musical coherence remains an important direction for future work.

ResponseMethod0.25σ0.5σ
TargetMask0.722.15
MWM3.4787.57
MWM+BHC3.0481.82
ReversalMask11.2914.38
MWM60.0397.67
MWM+BHC57.1297.34

3. Generalization.

MUSDB mainly provides audio data, while MWM works with symbolic music data such as MIDI, so it cannot be directly evaluated on MUSDB. Using this dataset would require converting audio into MIDI first, and errors introduced during this process would affect both generation quality and control performance, making it difficult to evaluate MWM itself. To better evaluate generalization across datasets and styles, we evaluated generation quality on additional MIDI datasets, including POP909 and Symphony MIDI (SymphonyNet), with in-domain Slakh-Action as a reference (below). We will further assess control performance and update our demo webpage with the corresponding results and generated music examples.

DatasetMethodRBDE↓C-JSD↓SC↑R-JSD↓GC↑RSE↓TSSE↓
Slakh-Action (in-domain)Executor-only0.4893.1759.8471.3678.6301.5320.891
MWM0.5673.2299.8471.3868.6521.5750.893
MWM+BHC0.4983.1739.8521.3658.6831.5300.890
POP909Executor-only0.5263.2229.8281.4058.5931.5740.913
MWM0.6043.2729.8351.4298.6141.6250.923
MWM+BHC0.5333.2199.8411.4008.6451.5660.909
Symphony MIDI (SymphonyNet)Executor-only0.5553.2879.8051.4588.5511.6290.949
MWM0.6343.3519.8091.4828.5781.6810.962
MWM+BHC0.5593.2759.8171.4498.6051.6160.945

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.