acceptodds
Under review as a conference paper at ICLR 2027

Kuramoto Decision Diffusion Models

Abstract

Diffusion-based decision making has seen explosive popularity due to the ability in producing rich, multimodal trajectories. However, existing frameworks erase the structure of robotic data, which is critical for action fluidity and non-myopic planning, through stochastic noising, then denoise or recover it only probabilistically. This paper aims to address the problem of incorporating spatio-temporal dependencies into the diffusion processes (i.e., forward and reverse), without sacrificing trajectory quality. Building upon and improving the work of Song et al., we propose KDDM, a physics-inspired diffusion paradigm that explicitly encodes the information through Kuramoto-based synchronization. By two neat tricks to harmonize the intertwined chains of Langevin and Kuramoto dynamics, we prove that KDDM elegantly converges to the desired outcome and theoretically outperforms prior approaches. Furthermore, we extend KDDM to ODE-accelerated and guided sampling, enabling its integration with today’s decision-making ecosystem. Experimentally, our proposals achieve superior performance in general, across an extensive range of tasks, including trajectory augmentation, long-horizon planning, and (real-world) robotic control, compared to various diffusion-based solutions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Intelligence, Diffusion Models, Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Yq6aOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper brings Kuramoto-style synchronization from KODM (Song et al., NeurIPS 2025) into diffusion models for decision making, used as planners, policies and trajectory synthesizers. Trajectory chunks are treated as phases of coupled oscillators. Instead of running stochastic Kuramoto dynamics, KDDM computes a deterministic mean-field coupling term from the clean chunk only (self-synchronization) and defines the forward kernel so that this term enters as a time-dependent mean shift, removed and re-added at every step (telescoping update). This gives closed-form Gaussian marginals and posteriors, a standard Gaussian prior and a clean-sample prediction loss. On top of this the paper derives an ODE sampler (KDIM), classifier and classifier-free guidance, replacement and posterior-sampling guidance, SMC, and a small network that approximates the coupling term. KDDM and KDIM are plugged into many existing pipelines: unconditional trajectory generation on MuJoCo and Adroit, D4RL planners built on Diffusion Veteran, Diffusion Policy, Diffusion BC and DSRL on Robomimic, PushT and Relay Kitchen, and four tasks on a real SO-ARM arm. Improvements are reported in temporal and spatial correlation of generated data, in D4RL averages, and most clearly for Diffusion BC.

KODM: Yue Song, T. Anderson Keller, Sevan Brodjian, Takeru Miyato, Yisong Yue, Pietro Perona, Max Welling; NeurIPS 2025, arXiv 2509.15328

Critical strengths
  1. The empirical scope is broad. KDDM is dropped into about ten existing methods across three roles (synthesizer, planner, policy) with small changes, and the guidance suite makes it usable in standard pipelines.
  2. The reformulation removes the practical problems of KODM in this setting (no closed-form marginal, no ODE sampler, no guidance), and as far as I could find this is the first use of Kuramoto dynamics in diffusion-based decision making. Training stays single-step like DDPM, apart from computing the coupling term.
  3. Results are consistent in direction: better temporal and spatial fidelity than DDPM-family baselines (Fig. 7), higher D4RL averages than the DV base (Table 1), and large gains for Diffusion BC, e.g. on PushT-Image (Fig. 9). The "w/o Syn." ablation on Adroit (Table 4) shows the shift helps over plain DDPM in the same code base.
  4. The paper is well-written and easy to follow.
Critical weaknesses
  1. The reviewer believes the technical novelty is overstated. Eqs. 4-6, the x0 objective and the DDIM version (Eq. 10) match ContextDiff (Yang et al., ICLR 2024) term for term, with its learned bias replaced by the Kuramoto term. ContextDiff is cited only in the appendix, while the intro says "We design" this kernel. ShiftDDPMs (Zhang et al., AAAI 2023) uses the same telescoping construction and Neural Diffusion Models (Bartosh et al., ICML 2024) covers it as a fixed-transform case; neither is cited. The only new parts are the Kuramoto coupling as the shift term and its use in decision making.
  2. It is unclear what the Kuramoto term encodes. KDDM is DDPM on the shifted variable, the prior is N(0, I), and the shift fades as synchronization completes. The coupling is global over all entries of the chunk, so it is blind to temporal adjacency and joint structure, against the spatio-temporal motivation. The phase mapping is not described, and a circular topology is odd for positions and velocities.
  3. The gains are not attributed. No ablation swaps the Kuramoto shift for another shift, and the DDPM baselines' parameterization is not stated while KDDM predicts x0. Diffusion Policy and DSRL gains are about one point, KDDM is within one or two points of PG on D4RL (PG is better on AntMaze), and Fig. 22 claims consistent planner gains though KDDM is below the original on several bars (its MuJoCo panels look transposed). The §5.2 numbers do not match Table 1, and the real-robot result is compared with 2/10 from another paper.

• Yang, L., Zhang, Z., Yu, Z., Liu, J., Xu, M., Ermon, S., and Cui, B. Cross-Modal Contextualized Diffusion Models for Text-Guided Visual Generation and Editing. ICLR 2024. arXiv:2402.16627. (ContextDiff) • Zhang, Z., Zhao, Z., Yu, J., and Tian, Q. ShiftDDPMs: Exploring Conditional Diffusion Models by Shifting Diffusion Trajectories. AAAI 2023. arXiv:2302.02373. • Bartosh, G., Vetrov, D., and Naesseth, C. A. Neural Diffusion Models. ICML 2024, PMLR 235. arXiv:2310.08337. (NDM)

Questions for authors
  1. What is technically new beyond the choice of coupling term? Eqs. 4-6 and 10 match ContextDiff's forward kernel, marginal, posterior and DDIM update with its learned bias replaced by κK, ShiftDDPMs has the same telescoping construction, and NDM contains KDDM as a fixed-transform case. Please cite these where Eq. 4 is introduced and restate the contribution relative to them.
  2. Can the gains be credited to Kuramoto dynamics? Please replace the Kuramoto shift with (a) a learned shift trained end to end as in ContextDiff, (b) a simple pull toward the chunk mean or a temporal smoothing operator, and (c) DDPM with an SNR-matched schedule, all with the same x0 parameterization and network. If (a) or (b) match KDDM on ACF/CC and on the planner and policy benchmarks, the gain comes from the shifted-diffusion construction rather than from synchronization.
7 comments
AuthorsOpenReview ↗

Official Response to Reviewer Yq6a (Part I)

Many thanks to you for providing thorough very thoughtful feedback. We have carefully reviewed all your comments and revised our manuscript. Due to space limits, the new and revised contents (marked in blue) have been integrated not only into the relevant sections of the main body, but also extensively into the appendix. We offer detailed responses to your comments below.

C1: Contributions over existing techniques.

    Response: We thank the reviewer for highlighting the connections to prior works, which have helped us better position our contributions. We have added citations to related works [1–4] at the appropriate places in the main text (e.g., Section 4.1). We acknowledge that the mathematical formulation of shifted diffusion itself is not entirely new. However, we would like to clarify three key distinctions:

  • Core contribution. The primary contribution of KDDM is not to propose another generic shifted diffusion formulation, but to introduce Kuramoto-based synchronization as a data-dependent mechanism for modeling spatiotemporal dependencies in robotic trajectories, and systematically develop this mechanism for diffusion-based decision making. This is also reflected in the narrative of our manuscript: KDDM is motivated by existing gaps (especially in KODM) in the decision-making literature, while the connection to shifted diffusion serves merely as a neat trick (as stated in Abstract) to address specific technical challenges. To avoid misunderstanding, we have revised the text in Introduction to emphasize our contributions.
  • Main differences. Despite sharing a similar high-level formulation, KDDM differs fundamentally from ContextDiff and ShiftDDPMs in several aspects. First, both are conditional generation frameworks that cannot be directly extended to our setting. Their terminal distributions, N(kTE(c,T),I)\mathcal{N}(k_T E(c, T), I), where EE denotes a learned shift predictor, depend on known conditions cc, making unconditional sampling infeasible. In contrast, thanks to the unique properties of Kuramoto dynamics (see Lemma 7), KDDM is fully self-contained and starts sampling directly from a standard Gaussian prior without requiring additional inputs. Second, KDDM offers greater interpretability and stability. The shift predictors in prior methods rely heavily on learned representations (e.g., CLIP), introducing substantial pretraining costs and potential numerical instability that may complicate the reconstruction of x0x_0 or ϵ\epsilon. In contrast, our Kuramoto system exhibits predictable synchronization dynamics with monotonically decreasing drift. Furthermore, while prior methods require assumptions about black-box model properties, such as Lipschitz continuity (lemma 4 in [2]), our framework naturally guarantees this property (see Lemma 6). To clarify these distinctions, we have added a subsection titled "In Comparison with ShiftDDPMs" in Appendix B.6. Regarding NDMs [3,4], which provide a general perspective on transformed diffusion processes, their target marginal distributions q(zt∣x)=N(αtF(x,t),σ2I)q(z_t|x) = \mathcal{N}(\alpha_t F(x, t), \sigma^2I) differ clearly from ours, so here we omit further discussion.
  • Border mechanisms. Although prior works also extend DDIM sampling, they do not provide the same comprehensive suite of conditional sampling mechanisms as KDDM, which supports Training-Based Guidance, Training-Free Guidance, and Sequential Monte Carlo (SMC) within a unified framework.

C2: Interpretation of Kuramoto coupling.

    Response: We would like to clarify that the purpose of Kuramoto coupling in KDDM is to model the collective coordination of trajectory elements rather than conventional smoothing. It is well-known that the Kuramoto model naturally incorporates spatio-temporal representations (e.g., see [5]). Following KODM, we treat the elements (both time and spatial axes) within a chunk, as coupled oscillators with the phase mapping θ=π×τ\theta = \pi \times \tau. A footnote has been added on Page 4 to describe the above. The circular geometry (also adopted from KODM) used here is standard in Kuramoto models. Their interactions are governed by relative phase differences, enabling the physics-inspired mechanism to capture collective synchronization patterns across the entire chunk, which can provide a useful inductive bias for trajectory generation. This allows trajectory elements to influence one another even when they are not adjacent; note that it is not blind to temporal adjacency and joint structure, because the ordering of all oscillators in the data vector (tensor) keeps intact. See demos in https://kuramoto-decision-diffusion.github.io/#background, in which the Phase Velocity is the encoding information we refer to. The underlying model in KDDM must perform denoising and desynchronization simultaneously upon receiving each θt\theta_t, therefore considering the spatiotemporal information.

AuthorsOpenReview ↗

Official Response to Reviewer Yq6a (Part II)

C3: Attribution of empirical improvements.

    Response: In response to the reviewer's request, in Section 5.5 we have conducted additional experiments (Data Synthesizer) comparing KDDM with alternative mechanisms under the same experimental setting. Specifically, we consider the following variants: (a) Learned shift: replacing the Kuramoto-based shift with a learnable shift mechanism using Ts2vec [6], a contrasive representation learning alogorithm. (b) Simple synchronization alternatives: using a mean-pulling mechanism to investigate whether the improvements can be explained by simpler forms of coordination. (c) Standard DDPM: using a comparable noise schedule and prediction parameterization to control for differences in diffusion configurations. All comparisons are designed to maintain consistent network architectures and prediction parameterizations wherever applicable. It can be seen in Figure 11, trajectories generated by (a) and (b) exhibit large performance degradation in modeling spatio-temporal dependencies (ACF and CC), with CC scores nearly collapsing, which we attribute to their numerical stochasticity. And (c) still underperforms KDDM. Moreover, as reported in the table below, for Diffusion BC in Relay Kitchen, we observe a similar trend that KDDM outperforms both shift alternatives on all four tasks and achieves the best overall performance. These results help distinguish the contribution of the general shifted-diffusion formulation from that of the specific synchronization mechanism.

MethodTask 1Task 2Task 3Task 4
KDDM100.092.777.745.4
(a) Learned shift28.670.670.00.0
(b) Mean Shift40.6725.05.330.0
(c) DDPM (x0)98.086.072.3348.33

C4: Numerical consistency and real-world validation.

    Response: We thank the reviewer for identifying several issues in the experimental presentation. First, we have also revised Figure 22 and corrected the disorder of the corresponding MuJoCo results. We apologize for these errors in the original manuscript. Second, regarding the real-world experiments, we have added direct comparisons against the baseline Diffusion Policy across all four real-world robotic manipulation tasks in Appendix D.6. We have also added some new demos of the baseline to https://kuramoto-decision-diffusion.github.io/#exp-robot. We believe these additional experiments provide stronger evidence for the practical effectiveness of KDDM.

TaskBaseline Success RateOur Success Rate
put-the-knife0.110.87
pick-the-bottle0.20.7
place-the-duck0.180.48
put-the-egg0.220.58

References:
    [1] Zhang, Zijian, et al. "Shiftddpms: Exploring conditional diffusion models by shifting diffusion trajectories." AAAI' 2023.
    [2] Yang, Ling, et al. "Cross-modal contextualized diffusion models for text-guided visual generation and editing." ICLR' 2024.
    [3] Nielsen, Beatrix MG, et al. "Diffenc: Variational diffusion with a learned encoder." ICLR' 2024.
    [4] Bartosh, Grigory, et al. "Neural diffusion models." ICML' 2024.
    [5] Miyato, Takeru, et al. "Artificial kuramoto oscillatory neurons." ICLR' 2025.
    [6] Yue, Zhihan, et al. "Ts2vec: Towards universal representation of time series." AAAI' 2022.

Dear Reviewer Yq6a,

We have incorporated the additional experiments and clarifications into our updated rebuttal. Whenever you have a moment to review it, we would greatly appreciate any further feedback. Thank you again for your time and consideration.

Best regards,
The Authors

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

This paper proposes Kuramoto Decision Diffusion Models (KDDM), which incorporate Kuramoto synchronization into diffusion models to better capture spatial and temporal dependencies in trajectories. The authors develop a new forward and reverse diffusion process with tractable distributions, and extend the method to accelerated and guided sampling. The paper includes theoretical analysis and extensive experiments on trajectory generation, offline reinforcement learning, imitation learning, and real-world robotic manipulation.

Critical strengths

The paper proposes an interesting approach to incorporating spatio-temporal dependencies into diffusion models. The theoretical development is substantial, and the proposed construction maintains useful properties of standard diffusion models, including tractable distributions and accelerated sampling.

The empirical evaluation is extensive, covering different decision-making problems and several diffusion-based methods. The reported results are generally positive, particularly for trajectory generation and imitation learning. The inclusion of real-world hardware experiments is also appreciated.

Critical weaknesses

The theoretical contributions and guarantees need further clarification. In particular, I am not fully convinced by the theoretical comparison with existing diffusion methods in Appendix B.6. Proposition 2 argues that local score matching is intrinsically suboptimal, but the variance of the local score target does not necessarily imply that the learned score function is suboptimal. The theoretical comparison with DDPM also relies on a simplified model and additional assumptions. It is not clear how generally these conclusions hold.

The empirical results could provide stronger evidence for the advantages of Kuramoto synchronization. While the results are generally positive, improvements over the strongest baselines are sometimes relatively small. The paper includes ablation studies on the synchronization mechanism, but I would like to better understand whether the benefits are specific to Kuramoto synchronization or could also be achieved using simpler ways of incorporating temporal and spatial dependencies. In addition, the real-world hardware experiments are rather limited. The authors report only 10 trials per task, without a direct comparison against standard diffusion policies under the same experimental setup.

Questions for authors

Could the authors clarify Proposition 2 and the theoretical comparison with standard diffusion methods? In particular, how does the variance of the local score-matching target establish that the learned score function is inferior?

Could the authors explain more clearly why Kuramoto synchronization is particularly suitable for robotic trajectories? Does the method improve physical or dynamical consistency, beyond improving statistical correlations? How would it compare with simpler structured diffusion approaches?

For the real-world experiments, could the authors provide direct comparisons with standard Diffusion Policy on the same robot, tasks, and demonstration datasets? This would help clarify the practical benefits of the proposed method.

2 comments
AuthorsOpenReview ↗

Official Response to Reviewer VaAp

We appreciate the reviewer’s recognition of our work’s main strengths: presenting a novel approach, delivering positive empirical results, and building on a solid theoretical foundation. We now address your concerns with additional experiments and the discussion below. All revisions in the manuscript are highlighted in blue.

C1: On the sub-optimality of Local Score Matching.

    Response: We thank the reviewer for pointing out this important issue. Proposition 2 shows that an individual local score generally deviates from the global score under non-degenerate posterior uncertainty and a non-vanishing Jacobian. It characterizes the intrinsic variance of local score targets. Here we acknowledge that this variance does not necessarily introduce asymptotic bias in the learned score function with exact posterior sampling and sufficient model capacity. However, the variance of local targets can increase gradient-estimation variance and make optimization less stable under finite-sample training (especially for high-quality trajectories). In contrast, our global score matching formulation directly supervises the global score, avoiding the need to estimate it through individual local transitions. Empirically, both Figure 7 (the NFID on some datasets) and Figure 13 (the convergence behavior of data points) support our interpretation. We have revised Proposition 2 and its remark to clarify that it establishes a discrepancy between local and global supervision signals, rather than an unconditional guarantee of inferior learned-score performance.

C2: Simplified theory model and additional assumptions.

    Response: We acknowledge that our theoretical comparison relies on a simplified Gaussian model and a linear approximation of the Kuramoto coupling. But we do not intend to claim that our advantage regarding the optimal Bayes estimation errors universally holds for arbitrary data distributions or nonlinear coupling dynamics. Instead, this analysis provides theoretical intuition for the potential benefits of synchronization in distribution modeling. Its practical relevance is further supported by our experimental results (e.g., better recovery of sptatial-temporal patterns). We have clarified the additional assumptions in Appendix B.1 and scope of this analysis to avoid overgeneralization.

C3: On the effect of Kuramoto-based synchronization.

    Response: We thank the reviewer for this insightful comment. First, we would like to emphasize that standard metrics (e.g., task completion rates) alone may not fully reflect the advantages of our proposal, as disscussed in [1]. For robotic trajectories, smoothness, naturalness, and temporal coherence are also important (see [2]). Kuramoto synchronization explicitly models cooperative interactions among trajectory chunk components, as illustrated by Figure 2, potentially enhancing these properties beyond what can be captured by statistical correlations alone. Therefore, we believe that the value of KDDM may be greater than what its performance metrics on RL and IL simulation tasks suggest, even though the results on these benchmarks are already encouraging overall.
    To better isolate the contribution of Kuramoto synchronization from that of the shifted construction itself, we have added two controlled comparisons in Section 5.5: (1) replacing the Kuramoto shift with a learned (representation) shift trained end-to-end; and (2) using simpler alternatives, including a direct pull toward the trajectory-chunk mean. Remarkable improvements from KDDM provides stronger evidence for the specific benefits of Kuramoto-based synchronization.

C4: Comparisons in the real-world experiment.

    Response: Following the reviewer's recommendation, we have added direct comparisons with the baseline Diffusion Policy across all four real-world tasks in Appendix D.6 (see also the table below, average over 10 times). One can see that KDDM consistently improves success rates (30–76 pp) and reduces completion times (20–61%), demonstrating its effectiveness in real-world manipulation. We have also updated our website (https://kuramoto-decision-diffusion.github.io/#exp-robot) to include the Baseline Diffusion Policy.

TaskBaseline SRBaseline TimeKDDM SRKDDM Time
put-the-knife0.1129s0.8719s
pick-the-bottle0.238s0.715s
place-the-duck0.1852s0.4830s
put-the-egg0.2220s0.5816s

References:
    [1] Higham, Sophie, et al. "Is Success All You Need? Investigating the Impact of Input Perturbations on VLA Behaviour in Tabletop Manipulation Tasks." arXiv preprint arXiv:2610.01351 (2026).
    [2] Liu, Chaoqi, et al. "Oat: Ordered action tokenization." RSS' 2026.

Dear Reviewer VaAp,

We hope this message finds you well. We wanted to let you know that the new experiments and clarifications have been added to our updated rebuttal. As less than two days remain in the Paper Discussion Phase, we would greatly appreciate your feedback. Thank you for your time and consideration.

Best regards,
The Authors

Reviewer_UgubOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

To address the challenge of incorporating spatio-temporal dependencies into diffusion processes, the authors propose KDDM, a novel approach based on Kuramoto synchronization. By introducing self-synchronization and telescoping updating techniques, KDDM can generate structured and high-quality trajectories. Extensive empirical and theoretical results are provided to demonstrate the effectiveness of the proposed method. Overall, this work is well-motivated and technically solid. Therefore, I recommend a weak accept at this stage, and I am willing to adjust my final rating based on the authors' rebuttal.

Critical strengths

The paper is well-organized, clearly written, and easy to follow.

The proposed KDDM achieves strong performance across a majority of the evaluated benchmarks.

The motivation behind the method is clear, compelling, and thoroughly discussed.

Critical weaknesses
  1. Potential degradation of trajectory stitching capability. A core mechanism of Kuramoto synchronization is to enhance the correlation of tuples within single trajectories from the dataset, while diminishing correlations with tuples from other trajectories (as suggested in Table 26). This could potentially degrade the agent's ability to "stitch" together sub-optimal trajectory segments into novel, optimal ones—a capability that is crucial for long-horizon navigation tasks in offline reinforcement learning.

  2. Lack of explicit presentation and justification of assumptions. Several underlying assumptions appear strong but are not explicitly formalized. Specifically:

  • 2.1) The framework seems to rely heavily on the assumption that the offline dataset is near-optimal to justify the behavior cloning approach. It remains unclear how "near-optimal" the dataset must be for the method to succeed, and how the sub-optimality of data affects the quality of the generated trajectories.

  • 2.2) In Line 1455, it is unclear why the equality holds for an arbitrary 1<t<T, especially considering that qσq_\sigma and qψq_\psi often differ in practice.

  • 2.3) In Line 1866, the assumption of a zero-error linear approximation for Kuramoto models seems impractical.

  • 2.4) Other mathematical and practical assumptions should be explicitly stated and discussed in a dedicated section or paragraph.

  1. Insufficient details regarding the encoder. The proposed method depends heavily on the encoder that maps trajectories to the latent oscillators θ\theta. However, technical details regarding this component are sparse. Is the encoder pre-trained or learned end-to-end? Furthermore, how do the authors theoretically or empirically guarantee that the learned θ\theta representations actually behave as valid oscillators within a Kuramoto system?
  2. Comparison with the PG baseline. The proposed KDDM and KDIM do not seem to consistently outperform PG (as shown in Table 1). The authors should provide a more nuanced discussion regarding the specific advantages of KDDM over PG. For instance, in what specific scenarios or under what data conditions is KDDM expected to yield superior performance?
Questions for authors
  1. Regarding trajectory stitching: Could the authors provide more insights or empirical evidence on how KDDM handles trajectory stitching? Specifically, how does the model perform on datasets that require stitching suboptimal trajectories to form an optimal one (e.g., "medium-replay" datasets)? Does the isolation effect of Kuramoto synchronization hinder this ability?

  2. Regarding dataset optimality: How sensitive is the proposed method to the quality of the offline dataset? I would appreciate it if the authors could clarify how KDDM performs when the dataset is highly sub-optimal and whether the behavior-cloning assumption is a strict requirement for the method's success.

  3. Clarification on mathematical derivations:

    • 3.1) Could the authors detail the derivation in Line 1455? Specifically, why does the equality hold for an arbitrary 1<t<T1 < t < T despite the potential divergence between qσq_\sigma and qψq_\psi?
    • 3.2) In Line 1866, what is the justification for the zero-error linear approximation? How would potential approximation errors propagate and affect the overall theoretical guarantees?
  4. Details on the encoder: Could the authors elaborate on the implementation and training pipeline of the encoder? Is it pre-trained or learned end-to-end? Furthermore, is there any empirical or theoretical validation to demonstrate that the learned latent representations θ\theta genuinely adhere to Kuramoto oscillator dynamics?

  5. Advantage over PG: Could the authors clarify the specific advantages of KDDM over PG? It would be helpful to discuss or demonstrate under what specific task structures, state-action spaces, or dataset properties KDDM is expected to demonstrate a clear superiority over PG.

4 comments
AuthorsOpenReview ↗

Official Response to Reviewer Ugub (Part I)

Thank the reviewer very much for your careful reading and positive feedback! Below we respond to your concerns, and our manuscript (marked in blue) has been revised accordingly.

C1: Potential degradation of trajectory stitching capability.

    Response: First of all, we would like to clarify that trajectory stitching is a critical capability in offline RL, but it is not the main focus of KDDM. Regarding the raised weakness, it indeed exists in all non-autoregressive diffusion paradigms, and enhancing intra-trajectory correlations does not necessarily eliminate the possibility of capturing useful patterns across different trajectories. A simple stitching solution (used in our work and Diffuser [1]) is to generate sufficiently long trajectories (covering all segments) and then perform inpainting. We presented its details and empirical evidence in Appendix D.1. Another way for long-horizon tasks requires designing additional conditioning models to mitigate the isolation effects, e.g., estimation-stitching-qualification workflow in [2] and bidirectional composition in [3]. Nonetheless, our Kuramoto synchronization can be directly integrated into these frameworks. To empirically validate this compatibility, we incorporate Kuramoto Drift into DiffStitch [2] and compare the resulting method against the original IQL baseline reported in the DiffStitch paper. As shown in the table below, KDDM achieves comparable, and in some cases even better, performance than the baselines reported in the original paper on the HalfCheetah environment, demonstrating its ability to stitch suboptimal trajectories into higher-quality ones.

Dataset NameOriginalDstitchDstitch+KDDM
Medium-Expert92.7±2.594.4±1.494.8±0.6
Medium48.5±0.349.4±0.1450.5±0.1
Medium-Replay44.1±0.544.7±0.144.9±0.3

C2: More details on the encoder.

    Response: In what follows, we answer the related questions one by one. (1) Implementation of the "encoder": By following [4], we simply rescaled the trajectory data τ\tau to the interval [−π,π][-\pi, \pi] during preprocessing as oscillators θ=π×τ\theta = \pi \times \tau. The (learned) encoder or mapping just takes the input t,θt, \theta and outputs the attractive force dθ/dtd\theta/dt among oscillators after tt steps of Kuramoto synchronization (as in line 230-232). Please see our random evolution demos in https://kuramoto-decision-diffusion.github.io/#background, in which the Phase Velocity is exactly the output of the "encoder" we refer to. (2) Pre-training pipeline: For the learned encoder, its pre-training is a standard supervised learning procedure that minimizes the reconstruction error between the NN's predictions and the GT phase velocities on the trajectory dataset. (3) Empirical/theoretical validation: We refer the reviewer to Figure 8 (left) in Section 5 and Figure 26 in Appendix E.1. The former demonstrates that the learned representations are well aligned with the Kuramoto system, while the latter visualizes the synchronization phenomena present in different trajectory datasets. Moreover, as evaluated in Appendix D.4, we show that the prediction error of the learned encoder almost always converges to below 0.01 (see Figure 23), and its performance in practice is comparable to that of the oracle encoder (see Figure 24).
    Correspondingly, we have added a footnote on Page 4 to clarify the mapping procedure, as well as a description of the learned encoder's architecture and training procedure in Appendix C.5.

C3: Advantage superiority over PG.

    Response: In essence, our proposed method is a plug-and-play diffusion paradigm. As discussed in the paper, KDDM and KDIM in Table 1 are two variants of DV,and the results generally demonstrate improvements in planning performance. Regarding PG, its underlying principle is very similar to DSRL [5], which is a post-tuning framework. It first trains a diffusion-based planner and then finetunes a prior sampling network. Thus, it is only suitable for offline reinforcement learning settings. This test-time approach trades additional computational time and resources for improved performance, which explains its relatively high D4RL scores. Of course, our KDDM can also be integrated into this framework. Due to rebuttal time constraints, we will consider including the corresponding experiments in Appendix D.3 (like Figure 22) in the next version.

AuthorsOpenReview ↗

Official Response to Reviewer Ugub (Part II)

C4: Lack of explicit presentation and justification of assumptions.

    Response: We appreciate the reviewer’s suggestions to clarify the underlying assumptions and provide more rigorous justifications for the mathematical derivations. Specifically,

  • About dataset optimality. We would like to clarify that the near-optimality assumption is not specific to KDDM; it is commonly adopted in many generative (control) solutions, as exemplified by [6]. We agree that the quality of the offline dataset can affect the performance of behavior cloning, but in fact, our goal is to enable KDDM to capture the spatiotemporal correlations in human demonstrations. Appropriate references have been added in the Problem Statement on Page 2.
  • About derivation of KDIM. Thanks for pointing out this ambiguity. The equality is proven through backward induction, rather than assumed a priori. Starting from the base case t=Tt=T, we derive the marginal distribution at step t−1t-1 using Eqn. (13). The drift terms cancel out in the mean, then the equality holds for all intermediate steps under the specified reverse transition.
  • About linear approximation. We clarify that this assumption is introduced to simplify the theoretical analysis by focusing on the overall effect of Kuramoto dynamics rather than explicitly modeling its nonlinear intermediate processes. Specifically, the coefficients of the state-dependent linear mapping capture the collective effects of all oscillators at only time tt. We acknowledge that this is an idealized approximation and have added its scope and in Appendix B.6.
  • About other assumptions. We have added a subsection of the underlying mathematical and practical assumptions in Appendix B.1, which we believe this will make the theoretical foundations and applicability of KDDM more transparent.

References:
    [1] Janner, Michael, et al. "Planning with diffusion for flexible behavior synthesis." ICML' 2022.
    [2] Li, Guanghe, et al. "Diffstitch: Boosting offline reinforcement learning with diffusion-based trajectory stitching." ICML' 2024.
    [3] Luo, Yunhao, et al. "Generative trajectory stitching through diffusion composition." NeurIPS' 2026.
    [4] Song, Yue, et al. "Kuramoto orientation diffusion models." NeurIPS' 2025.
    [5] Wagenmaker, Andrew, et al. "Steering your diffusion policy with latent space reinforcement learning." CoRL' 2025.
    [6] Hong, Matthew M., et al. "TMRL: Diffusion Timestep-Modulated Pretraining Enables Exploration for Efficient Policy Finetuning." RSS' 2026.

Thank you for the detailed response. My main concerns have been solved basically, so I will maintain my original positive rating and raise my confidence score.

Thanks for your comment! We are glad to have addressed your main concerns and would be happy to clarify any remaining questions or finer details.

Comments on the paper

2 comments

Dear reviewers,

The authors have now submitted their responses. Please review them and share your thoughts on whether your concerns have been addressed. I encourage you to engage with one another’s assessments, discuss any remaining disagreements, and update your reviews and scores as appropriate.

Thank you.

AC

AuthorsOpenReview ↗

General Responses and Summary of Revisions

We are grateful to the reviewers (R1=Ugub, R2=VaAp, R3=Yq6a) for their careful reading and constructive feedback. We are pleased that all reviewers recognized the contribution of our work: introducing Kuramoto-based synchronization into diffusion-based decision making is novel and, as far as we know (and as R3 independently verified), the first use of Kuramoto dynamics in this area; R1 further found the motivation clear and the method technically solid. Most reviewers also found the work interesting, well-written, and easy to follow, and appreciated the broad empirical scope including the real-world manipulation experiments, as well as the theoretical development, the tractable distributions and accelerated sampling, and the plug-and-play nature of KDDM.

We summarize major changes in the revision as follows.

  • Real-World Comparisons (R2, R3). Added comparisons with the Diffusion Policy baseline on all four real-world tasks (Appendix D.6) under the same platform, tasks, and demonstration datasets (over 10 trials per task). Baseline demos are now on our project page.
  • Attribution of Gains (R1, R2, R3). To separate our Kuramoto model from the shifted-diffusion construction, Section 5.5 now reports controlled comparisons under the same network and x0x_0 parameterization: (a) a learned end-to-end shift, (b) a mean-pulling shift toward the chunk mean, and (c) a setting-matched DDPM.
  • Contributions over Prior Work (R3). Added appropriate citations (Section 4.1), revised the Introduction, and added "In Comparison with ShiftDDPMs" (Appendix B.6).
  • Clarifications of Assumptions/Derivations (R1, R2, R3). Revised Proposition 2 and its remark to describe a discrepancy between local SM and global SM; consolidated all assumptions in Appendix B.1 and scoped the comparison in Appendix B.6.
  • Corrections and More Details (R1, R2, R3). Added a footnote on the phase mapping (Page 4), documented the encoder in Appendix C.5, and corrected the MuJoCo panel ordering in Figure 22.

All updates are now highlighted in blue. Compared with the original submissions, the revised manuscript has additional 3 pages (52 →\rightarrow 55).

The valuable suggestions from reviewers are very helpful for us to revise the paper to a better shape. We would be glad to answer any further questions. Looking forward to the reviewer's feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.