acceptodds
Withdrawn from ICLR 2027

MGPO: Manifold-Guided Diffusion Alignment for Task-Aware Dataset Distillation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion-based dataset distillation (DD) suffers from a fundamental objective mismatch: likelihood-driven diffusion models prioritize density approximation over the discriminative decision boundaries required for downstream tasks. Beyond semantic mismatch, relying solely on density also leads to geometric coverage loss, where generated samples collapse into a few high-density modes and fail to cover the manifold’s structural diversity. We propose Manifold-Guided Policy Optimization (), which reformulates DD as a multi-objective reinforcement learning problem and achieves Dual-Space Alignment via a pixel-space discrimina- tive reward and a latent-space geometric reward guided by a class-wise Minimum Spanning Tree (MST). The discriminative reward enforces class separability, while the MST-based geometric reward encourages generated latents to cover a sparse geometric skeleton of each class, jointly addressing both failure modes. We further provide an idealized analysis that motivates the MST-based reward, including a Hausdorff approximation bound and a subsampling bound independent of the dataset size. The reward-modular design extends to structured tasks such as object detection and segmentation by substituting the frozen task reward model. Extensive experiments show consistently outperforms existing methods, including a +8.0% mIoU gain on segmentation under low-budget setting

Primary area:
generative models
Keywords:
Dataset Distillation, Reinforcement Learning

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_gPj9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

MGPO trains a diffusion model to create small synthetic training sets. It uses reinforcement learning with two rewards. A fixed task model gives the first reward. This reward measures whether an image matches the requested class or task. The second reward uses a minimum spanning tree (MST) built from real image features. It rewards generated features that are close to real nodes. It also rewards generation steps that follow nearby tree edges. The paper tests classification, object detection, and segmentation. It also gives theoretical results about how well the tree approximates the data geometry.

Critical strengths
  1. The problem has practical value. Perceptual quality is not equivalent to the training value of a small data set; directly considering the task signal in the generation stage is a reasonable route. Replacing offline optimization costs with a generation process that does not require reward guidance also has potential value.
Critical weaknesses
  1. The paper gives two different direction rewards. Equation (4), on page 5, uses 1 - |cos|. Appendix C.2, on page 25, uses 1 - cos, without the absolute value. I checked both formulas in the rendered PDF. The difference changes the reward. An MST edge has no fixed direction. Reversing its endpoints changes the signed version, but leaves the absolute version unchanged. The two penalties also have different ranges: [0,2] and [0,1]. Appendix A uses the absolute version. It does not establish which version produced the results. The authors must identify the actual implementation and make the formulas consistent. If they used the signed version, they should test whether reversing edges changes the results.
  2. Two parts of the theory need correction. Theorem 10 gives a sample threshold that does not depend on the confidence parameter η. It requires C(log(n/η)/n)^(1/m) to stay below a fixed geometric threshold. For fixed n, this expression grows without limit as η approaches zero. Thus, the stated result cannot hold for every η in (0,1). Corollary 13 uses an excess-risk theorem that the paper does not provide. This missing theorem should connect the generated data to the classifier's test risk. I searched the complete PDF and checked Appendix A. The required theorem and assumptions are missing. Theorem 10 also uses edge distance where Equation (4) uses node distance. Its reward expression omits the distance coefficient from Equation (4). The authors need consistent definitions and a complete proof for the risk claim.
Questions for authors

See weakness.

Reviewer_y6daOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Manifold-Guided Policy Optimization (MGPO), a reinforcement learning framework for task-aware diffusion-based dataset distillation. The method addresses the mismatch between likelihood-driven image generation and downstream discriminative objectives, as well as insufficient geometric coverage of the data manifold. Specifically, MGPO introduces a dual-space reward mechanism that combines a pixel-space discriminative reward with a latent-space geometric reward based on class-wise Minimum Spanning Trees (MSTs). These rewards are incorporated into group-wise policy optimization to fine-tune pretrained diffusion models without requiring inference-time guidance. The authors also provide theoretical analyses of MST-based manifold approximation and subsampling scalability. Extensive experiments on image classification, object detection, and semantic segmentation benchmarks demonstrate the effectiveness of the proposed framework, with reported improvements over several existing dataset distillation methods.

Critical strengths
  1. The paper addresses the mismatch between likelihood-driven generation and downstream learning objectives by combining discriminative and geometric rewards within a reinforcement learning framework. This formulation is interesting because it provides a flexible way to incorporate task-specific objectives into the diffusion model without requiring additional inference-time guidance.

  2. The paper evaluates MGPO on several classification benchmarks, including ImageNet-1K, and further extends the framework to object detection and semantic segmentation. The reported improvements, particularly in low-budget segmentation settings, suggest that the approach has practical potential beyond standard classification-based dataset distillation.

Critical weaknesses
  1. The paper's central claim is that the proposed geometric reward improves manifold coverage and mitigates mode collapse. However, the reward primarily minimizes the distance between individual generated samples and their nearest MST nodes, without explicitly encouraging coverage of different manifold regions. Multiple samples could therefore concentrate around the same node while receiving high rewards. Moreover, the distance component does not depend on the MST edges, raising questions about the necessity of the proposed geometric structure. This concern directly affects the validity and novelty of the paper's main contribution.
  2. The theoretical analysis does not fully correspond to the implemented method, particularly regarding the use of nearest-node distances, PCA representations, and directional alignment. Several mathematical statements also require clarification, including the assumptions in Theorem 10 and the excess-risk bound referenced in Corollary 13. Meanwhile, the improvements over strong baselines are sometimes modest despite the additional RL fine-tuning and geometric preprocessing. Without stronger theoretical consistency, targeted ablations, and compute-matched comparisons, it is difficult to determine whether the proposed complexity is sufficiently justified by the reported gains.
Questions for authors
  1. How does the proposed nearest-node distance reward explicitly encourage manifold coverage rather than simply promoting proximity to reference samples? Since multiple generated samples could concentrate around the same MST node while receiving high rewards, what mechanism prevents such behavior? Additionally, could the authors provide comparisons against simpler nearest-neighbor or kNN-based regularization to demonstrate that the MST structure itself is responsible for the improvements? Convincing evidence on this point would substantially strengthen my assessment of the method's novelty and effectiveness.

  2. Clarification on theoretical validity and implementation consistency. Could the authors clarify the relationship between the theoretical guarantees and the implemented geometric reward, particularly given the differences in distance definitions and latent representations? Specifically, Theorem 10 appears to require a sample-size threshold that depends on the confidence parameter, while Corollary 13 refers to an excess-risk bound that is not established in the main text. Furthermore, which directional alignment formulation is actually implemented: the absolute cosine similarity in Eq. (4) or the signed inner product in Appendix C.2? Resolving these issues would materially affect my assessment of the paper's technical soundness.

  3. Could the authors report the total computational cost of MGPO, including RL fine-tuning, reward computation, and MST preprocessing, and compare it with strong baselines under matched computational budgets? Given the relatively modest improvements on ImageNet-1K and several ImageWoof settings, understanding whether the performance gains justify the additional optimization complexity would significantly influence my overall recommendation.

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

MGPO formulates diffusion-based dataset distillation as reward-driven policy optimization. It combines a pixel-space task reward with a latent-space geometric reward constructed from class-wise minimum spanning trees. Group-wise normalization combines the rewards during diffusion-model fine-tuning. The framework is evaluated on classification, detection, and segmentation, with structured annotations obtained through frozen task-model pseudo-labeling.

Critical strengths

Aligning a generative model with downstream training utility is a relevant and well-motivated objective. The reward-modular formulation offers a common optimization interface across several task families. Classification experiments cover multiple datasets and student architectures, and the reported segmentation improvement over UniDD is substantial at the lowest budget. The pixel/latent reward ablation provides evidence that the two signals can complement one another.

Critical weaknesses
  • The necessity and effectiveness of the MST construction are insufficiently isolated. Nearest-node attraction and local directional alignment do not themselves require generated samples to cover different modes. Table 6 compares pixel and latent rewards, but does not distinguish MST-specific benefits from generic reference-node proximity or simpler graph regularization. Quantitative coverage diagnostics and matched node-only/kNN/random-tree controls would directly test the central geometric claim.

  • The structured-task rewards are imperfect proxies for the claimed spatial utility. Maximum detection confidence does not directly assess localization, while top-K segmentation confidence does not measure mask completeness or boundary accuracy. Since the same frozen model also supplies pseudo-labels, the experiments need to separate the benefit of policy alignment from the quality and biases of the teacher-generated annotations. The explicitly single-class formulation also limits the scope of the structured-task generality claims.

Questions for authors
  1. Under matched reference nodes and optimization budgets, how does the MST reward compare with node-distance-only, kNN, and random-tree alternatives? Do coverage measurements support a benefit beyond proximity to real samples?

  2. How well do the structured-task rewards correlate with localization or mask quality, and how does MGPO compare with a matched frozen-generator plus pseudo-labeling/filtering baseline?

  3. Where is the excess-risk result required by Corollary 13 established?

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

Related papers

Discussion (0)

Sign in to comment.