acceptodds
Under review as a conference paper at ICLR 2027

Quotients Beyond Group Actions: Learning Variable Equivalence with Local Geometry

Abstract

Learning equivalence classes removes redundancy by identifying different realizations that represent the same input. Existing group-based equivalence methods accommodate redundant variation through a shared transformation structure, using the group action to define both the equivalence classes and their quotient geometry. This restricts the resulting equivalence to classes generated by that action, limiting task-, context-, and realization-dependent or non-geometric equivalence. We propose a novel approach to representation learning that learns a quotient map directly from observed outcomes, instead of deriving it from predefined transformations, by assigning shared coordinates to equivalent realizations. In continuous domains, the learned quotient map realizes equivalence classes as smooth level sets, with its differential and the Riemannian metric defining the tangent spaces and orthogonal decomposition, yielding a local quotient geometry. We use this geometry with flow matching to generate preferred realizations within fixed equivalence classes and to define consistent motion across changing classes. Experiments show that the learned quotient captures equivalence across both image and language domains, while articulated system experiments demonstrate generation and dynamics on the learned quotient geometry.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Equivalence Learning, Representation Learning, Quotient Space, Constrained Generative Modeling, Invariance and Equivariance

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3G6GOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Representing data as equivalence classes is effective in reducing redundancy; however, existing group-based works are restricted by the definition of group actions. Thus, such previous works cannot be applied to tasks where defining a group action is difficult, such as natural language. The authors propose a novel approach to identify equivalence classes and construct a quotient space directly from observations rather than relying on predefined transformations. Then, a conditional flow matching model generates realizations along the horizontal and vertical directions of the fiber.

Critical strengths
  1. I think that constructing a quotient space without explicitly defining a group is, ironically, a promising way to implement the effects of group actions. In many cases, we do not require the explicit group representation itself. Thus, I think this work provides a solid starting point for such a research direction.

  2. The theoretical setup and results, as well as the implementation with flow matching, are solid.

  3. The empirical evidence is sufficiently broad and shows clear improvements.

Critical weaknesses
  1. In the second paragraph of Section 3.3, the existence of a "constant rank rr" is not guaranteed. The authors commented on this in the Appendix; however, this issue should be treated more carefully because the proposed method relies on the constant-rank theorem. I think that stating an explicit assumption would alleviate this issue.
  2. In Table 3 and Table 4, the SR is noticeably low except for the 3 Joint + obs case.
Questions for authors
  1. Regarding W1: Can you show the rank stability and the singular value distribution on actual data, even in a simple case?

  2. How do approximation errors in the learned quotient map affect the topology and geometry of the resulting fibers?

  3. Regarding W2: Is there a trade-off between SR and SW?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3G6G

Thank you for your insightful review and encouraging assessment. We appreciate your recognition of our theoretical contributions, flow-matching implementation, and broad empirical validation. Your questions regarding rank stability, quotient-map approximation errors, and the trade-off between success rate and distributional fidelity are particularly valuable. Below, we address each point in detail.

 


  • W1&Q1 (Constant-rank assumption and empirical rank stability):

We agree that the constant-rank assumption should be explicit and have revised Sec. 3.3 to explicitly restrict each local decoder to a C2C^2, constant-rank region UiU_i, without crossing rank-changing boundaries. Rank may vary across regions, including ri=kr_i=k (zero-dimensional fibers). To empirically assess this assumption, we evaluated 6,000 held-out Panda configurations across three seeds. All exhibited learned rank 6 at threshold 10−710^{-7}, with minimum σ6=3.60×10−4\sigma_6=3.60\times10^{-4} (Sec. 4.3). The learned singular-value distributions closely match analytical kinematics.

Singular valueAnalyticalLearned
σ₁5.061 [4.488, 5.658]5.061 [4.489, 5.658]
σ₂3.655 [2.929, 4.289]3.655 [2.930, 4.290]
σ₃2.842 [2.060, 3.456]2.842 [2.059, 3.454]
σ₄0.888 [0.577, 1.253]0.888 [0.577, 1.253]
σ₅0.555 [0.316, 0.906]0.555 [0.315, 0.906]
σ₆0.189 [0.017, 0.372]0.189 [0.017, 0.372]
Minimum σ₆1.29 × 10⁻³3.60 × 10⁻⁴

We further tested local rank stability using 16 directional perturbations per center at varying normalized joint-space radii.

RadiusStable / in-boundsRankStability
10⁻⁴768/7686100%
10⁻³768/7686100%
10⁻²735/7356100%
0.05606/6066100%

These results support local rank stability within the tested neighborhoods, without claiming globally constant rank or guaranteeing regularity at untested points.

 


  • Q2 (Effects of approximation errors on fiber geometry and topology):

Approximation errors in the learned quotient map can shift fibers and misalign their tangent spaces. Near singularities, they may also alter fiber connectivity or dimension, so global topology preservation is not guaranteed. Empirically, the learned Panda nullspace closely matches analytical geometry, with a median angular error of 0.61°, mean of 1.61°, and 95th percentile of 5.83°. We further evaluate physical position error under finite displacement using identical learned geometry.

Displacement / chart radiusLinear tangent (mm)Implicit decoder (mm)
0.000.6270.627
0.251.0160.627
0.552.9220.633
0.704.4300.638

At 70% of the chart radius, implicit decoding reduces physical error from 4.430 mm to 0.638 mm, remaining close to the initial 0.627 mm error (Fig. 7, App. C.4). This demonstrates that the nonlinear correction limits additional drift despite approximation errors, supporting local geometric fidelity without claiming global topological guarantees.

 


  • W2&Q3 (Success rate and Distribution accuracy):

We clarify that our method achieves 99.99% SR on 3J, 97.09% on 3J with obstacles, and 95.18% on Panda, with a lower 75.13% on 4J. Where ours is not the highest, the SR gap is at most 0.85 points, while our SW₁ is the lowest in all settings. Notably, all projection-based baselines exhibit similarly reduced 4J success rates (75.55–75.98%), suggesting a shared experimental difficulty rather than a limitation specific to our method. This is associated with the learned outcome map, whose mean endpoint error of 0.754 mm is close to the strict 1 mm tolerance, and exact geometry restores 100% success for both ours and FM with final projection.

Regarding the trade-off, SR measures task feasibility and target satisfaction, whereas SW₁ measures distribution accuracy. There is no inherent trade-off between them. In the matched exact-geometry ablation (Tab. 14), all variants achieve 100% validity, while SW₁ and component TV improve substantially. Similarly, in Tab. 13, all local-geometry baselines achieve 100% validity, while ours achieves the lowest SW₁ and component TV. Tab. 14 is attached below for reference:

VariantValid (%) ↑SW₁ ↓Component TV ↓
w/o coarea1000.044240.08051
w/o volume1000.099610.19085
Ours1000.015050.02844

These results demonstrate that distribution fidelity can improve without sacrificing feasibility, rather than implying an unavoidable trade-off between SR and SW₁.

Thank you for the detailed rebuttal and additional experiments. My main concerns have been addressed.

One additional question came to mind regarding computational cost. Have you measured the computational overhead of the proposed method compared to the baselines, particularly in terms of training and inference time? If so, a brief explanation would be helpful.

AuthorsOpenReview ↗

Response to Official Comment by Reviewer 3G6G

Thank you for this question, and we are glad that all the main concerns have been addressed.

We measured both the additional geometry/preparation cost and complete sampling cost of the heaviest experiment, which is the spatial arm experiment, using RTX 5090, Xeon Silver 4510, CUDA-synchronized timing. Our generator is small and cheap to train, but the method adds offline preparation and a one-time local decoder construction per new target pose (~239 s). The extra computation buys fiber preservation and control.

Inference vs native baselines (2048 output)

MethodPer-target setup (s)Sampling (s)SR (%)
IKFlow + native refinement–0.3293.87
GGIK–0.130.03
Ours238.560.8097.95

Setup is performed once per target and reused across batches. The native baselines are run without our terminal projection, which explains GGIK's low success (with projection (Tab. 12), GGIK reaches 99.35%).

Training vs native baselines

ComponentOursIKFlowGGIK
Outcome-map training20.2 min––
Training-bank preparation48.4 h (summed over parallel jobs)released datareleased data
Generator training7.0 minnot measurednot measured
Forward/backward, batch 5122.19 ms21.75 ms152.77 ms

The outcome map and banks are shared with the learned geometry baselines in Tab. 4, but are additional cost relative to IKFlow and GGIK. Bank preparation constructs the cost weighted training configurations on each fiber and is run as parallel jobs, and the 48.4h is the total duration of the bank preparation (12.1h on 4 GPUs).

Methods sharing our local geometry

MethodGenerator training (min)Sampling (s)
RFM (adapted)175.963.99
PCFM (adapted)24.137.06
Ours7.010.80

RFM and PCFM here uses our local geometry and decoder, so preparation costs are shared. Ours learns a scalar local coordinate and decodes once, while RFM builds numerical geodesic training paths and PCFM repeatedly corrects configurations during sampling.

Reviewer_WHj2OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a novel framework for learning behavior-defined equivalence relations beyond predefined group actions. Rather than relying on a shared global transformation structure, the proposed method learns quotient representations directly from task-relevant outcomes, allowing equivalence to depend on the task, context, and individual realization.​For continuous domains, the authors leverage the differential of the learned quotient map and Riemannian geometry to characterize local equivalence classes and their associated geometric structure. Building upon this formulation, they further develop fiber-preserving generation and projectable dynamics through implicit decoding and flow matching, enabling the generation of different realizations that satisfy the same task-relevant constraints.​The authors conduct extensive experiments across multiple domains, including image-based equivalence recognition and local geometry recovery, language-based semantic equivalence, and constrained generation and dynamics for planar and spatial robotic arms. These experiments provide empirical evidence supporting different aspects of the proposed framework.

Critical strengths

1. Strong methodological novelty with a solid theoretical foundation.

The paper presents an innovative framework that extends equivalence learning beyond predefined group actions. It establishes a principled connection between quotient representation learning, local geometry, and constrained generative modeling. Importantly, translating this general theoretical framework into working algorithms is itself a significant contribution.

2. Comprehensive and progressively structured experimental validation.

The experiments are well organized and progressively validate different aspects of the framework, from equivalence recognition and local geometry recovery to language equivalence and constrained generation and dynamics in robotic systems. The extensive comparisons, ablations, and analyses provide a comprehensive initial evaluation, which is particularly valuable for a newly proposed theoretical framework.

3. Promising empirical results demonstrating the feasibility of the framework.

The method achieves substantial improvements in local geometric recovery and demonstrates strong performance in controlled language equivalence and robotic generation tasks. These results provide encouraging evidence that the proposed theoretical framework can be successfully implemented and produce meaningful practical benefits.

Critical weaknesses

1. Limited validation on representative real-world applications.

Although the experiments comprehensively evaluate the theoretical properties of the framework, they are primarily conducted in controlled or simulated settings. A more representative and practically meaningful application demonstrating clear advantages in task performance, generalization, or robustness would substantially strengthen the paper.

2. Insufficient comparisons with established methods on mainstream benchmarks.

While the paper includes several relevant baselines, most comparisons are conducted within the authors' selected experimental settings. More extensive evaluations on widely adopted benchmarks against strong established methods would better demonstrate the practical significance and advantages of the proposed framework.

3. Continued reliance on manually defined equivalence relations.

Although eliminating the need for predefined group actions is a significant advancement, the proposed framework still relies on manually specified task-relevant behavior to define equivalence classes. This dependence may limit its generality and applicability to tasks where the appropriate equivalence relations are unknown or difficult to specify.

Questions for authors

Q1. Practical applications and real-world relevance.

The proposed framework is theoretically interesting and has been evaluated through a comprehensive set of controlled experiments. Could the authors further demonstrate its effectiveness on a more representative real-world task or a widely adopted benchmark? In particular, it would be valuable to identify a practical scenario where behavior-defined equivalence provides clear benefits in terms of task performance, generalization, or robustness.

Q2. Comparisons with established methods.

Could the authors provide more extensive comparisons with strong, established methods on mainstream benchmarks? In particular, under what conditions does the proposed framework offer distinct advantages over conventional representation learning, group-based approaches, or existing constrained generative methods? Such comparisons would help clarify the practical significance of the proposed framework.

Q3. Generality and dependence on predefined equivalence relations.

Although the framework removes the need for predefined group actions, it still relies on manually specified task-relevant behavior to define equivalence classes. How sensitive is the method to the choice and quality of these behavior definitions? Could the framework be extended to settings where equivalence relations are unknown, only partially specified, or must be inferred from data?

1 comment
AuthorsOpenReview ↗

Response to Reviewer WHj2

Thank you for your thoughtful and comprehensive review. We are encouraged by your recognition of our methodological novelty, theoretical foundations, and extensive experimental validation. We also appreciate your valuable suggestions regarding real-world applicability, stronger baseline comparisons, and the generality of behavior-defined equivalence. Below, we address your questions and concerns in detail.

 


  • W1&Q1 (Practical applications and real-world relevance):

We thank the reviewer for emphasizing practical validation. We additionally evaluate our method on the established RLBench robotics manipulation benchmark, including drawer and microwave opening tasks. Across three seeds, six held-out scenes, and 180 requests per method, ours achieves 88.33% vs. 85.00% opening success on drawers and 99.44% vs. 96.11% on microwaves compared with standard FM. Our anonymous robot experiment page:

https://anonymous.4open.science/r/ICLR_Rebuttal-36184/01_rlbench/README.md

shows exploration from expert grasps ("Exploration from Expert Demo"), and diverse generated configurations with successful opening executions ("Grasp Variations in Learned Vertical Direction"). These experiments extend our evaluation to task-level robotic manipulation and held-out scene generalization. Beyond this, as noted in the Conclusion, implementation in more complex tasks, including physical robot deployment, remains an important direction for future work.

 


  • W2&Q2 (Comparison with established methods):

Following your suggestion, we evaluated our method on RLBench, as mentioned above, a recognized robotic manipulation benchmark, against standard FM under the same execution controller. Ours improves opening success from 85.00% to 88.33% on drawers and from 96.11% to 99.44% on microwaves. Beyond these success rates, our anonymous robot experiment page visualizes exploration from a single expert demonstration, the learned task-equivalence fibers with their vertical and horizontal directions, and diverse grasps along the learned vertical direction that successfully open both the drawer and the microwave. These results complement the established baselines in our manuscript, including representation-learning methods such as STN, SupCon, RNC, and SFA, and generative and constrained methods such as RFM, PCFM, IKFlow, and GGIK. Our framework is most advantageous when equivalence is not generated by a group action (Tab. 2), when the distribution over equivalent realizations matters (about 80% lower SW₁ and component TV on Panda, Tab. 4), and when local geometry must be recovered (82.13% vs. at most 51.41% rank accuracy, Tab. 1).

 


  • W3&Q3 (Generality and dependence on predefined equivalence relations):

Our framework requires specifying which task outcome to preserve, not which transformations preserve it. In the new RLBench experiments, we perturb demonstrated 6D grasp poses and observe their opening behavior through simulation rollouts. The learned behavior map identifies locally equivalent grasps without prescribed symmetry groups or known invariant transformations. The robot experiment page visualizes the learned vertical and horizontal directions ("Creating Task-equivalence Fiber") and verified alternative grasps ("Grasp Variations in Learned Vertical Direction"). Thus, equivalence-preserving variations are inferred from observed outcomes rather than manually specified. Regarding sensitivity to behavior quality, our experiments already use approximate behavior, from a finite panel of 32 worlds in language (App. C.2) and a learned outcome map for the arms, where ours achieves the lowest distribution error with both learned and exact geometry (Tab. 4, Tab. 13). In RLBench, the behavior map itself is learned from rollouts. Where no continuous local redundancy exists, the framework also permits zero-dimensional fibers.

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

This paper introduces a framework of learning equivalence directly from data outcomes rather than predefined transformations. To achieve this goal, the authors develop a behavior-defined quotient representation, a cost-sensitive distribution over a fiber sampled by flow matching, and a projectable dynamics that makes equivalent realizations follow the same quotient trajectory. To validate the proposed method, the authors design experiments that cover context-dependent equivalence on MNIST (relies on encoder-Jacobian supervision), CFQ-derived language equivalence, and robotics tasks, where the method reports lower distribution error for fixed-target generation (planar/spatial arm) and higher execution success for moving-target tracking (Panda).

Critical strengths

-The idea of learning equivalence directly from data outcomes rather than predefined transformations is new and interesting.

-The problem seems to be well formulated and mathematically solid.

Critical weaknesses

-The central claim of “learned equivalence beyond group actions” is weakly tested. The experimental settings are either synthetic with known transformations, or it is not clear if an unknown equivalence structure is identified from the proposed method.

-Many of the notations in the appendix do not match with the main paper. The authors shall carefully check and make them consistent to remove confusions. For example, in appendix I.2, the authors mentioned training the encoder EcE_c, however in the main paper there is no EcE_c or encoder.

-The main document should be self-sufficient for a reader to fully understand the paper. Currently, as the manuscript is written, a reader needs to go to the appendix repeatedly to understand the method. The vast majority of the important details are actually pushed into the appendix which should be in the main paper.

Questions for authors

Thanks for the authors' efforts addressing my questions below. I am happy to increase my score.

-What is the computational cost and time complexity of the proposed method? It looks like it requires SVD to calculate the orthogonal basis EiVE^V_i and EiHE^H_i in equation (10). The eigenvalue decomposition requires a large amount of computational resources if the data dimension is high.

-Moreover, why is there an additional correction term hih_i is needed or enforced in Eq (10)? Such important information should be included in the main text rather than Appendix.

-How sensitive are the vertical/horizontal split and downstream generation to the singular-value threshold?

-The loss function to train the quotient map qcq_c seems difficult to optimize. For example, in the MNIST experiment, the overall loss contains a lot of individual loss terms zipped together with weighting parameters (equation I.4, page 21). What is the motivation for using some of these loss terms such as L_id, L_comp, L_coord, and etc.? How are the weighting terms determined?

-For MNIST dataset, is it possible to apply the proposed method for hand-written digit generation task? Is it possible to generate different transformations of the same digit using their flow matching algorithm where the generated digits are within the same fiber?

3 comments
AuthorsOpenReview ↗

Response to Reviewer muxE (1/3)

Thank you for your careful review and constructive feedback. We appreciate your recognition of the novelty and mathematical soundness of our framework. We also value your suggestions regarding experimental validation, computational efficiency, numerical sensitivity, and presentation clarity. Below, we address each of your concerns in detail.

 


  • W1 (Learning beyond group actions):

Regarding the central claim of “learned equivalence beyond group actions”, we intended “beyond group actions” to mean beyond explicitly specified group actions, where “explicit” was omitted for brevity. We agree this was unclear, and we will check whether the title can be edited after submission and revise it accordingly if possible. The Introduction describes prior methods as requiring a global transformation group “known in advance” (p. 1) and notes that task-dependent equivalences may lack a “single known group action that is globally valid” (p. 1). The abstract contrasts our approach with “predefined transformations.” We do not claim the equivalences cannot be represented by any group action, only that no such action must be supplied. Our language experiment (Sec. 4.2) already shows this, as equivalent questions share no transformation group and are related only through answer behavior.

Additionally, to back up our claim, we introduce robotic manipulation experiments that learn which grasp variations preserve task success from observed outcomes, without prescribing invariant transformations. Specifically, we evaluate grasp generation using the RLBench [1] dataset for opening drawers and microwaves. Starting from a single expert demo, we perturb their 6D poses, execute grasp-and-open attempts in simulation, and learn a behavior map predicting opening progress at four stages. Importantly, these perturbations are exploratory interventions, not transformations assumed to preserve equivalence. The learned map identifies locally behavior-preserving directions, which our implicit decoder and flow matching use to generate alternative grasps. Both our method and standard FM use the same execution controller.

Across 3 seeds, 6 held-out scenes, and 180 requests per method, ours achieves 88.33% vs. 85.00% opening success on drawers and 99.44% vs. 96.11% on microwaves (ours vs. standard FM, SD 3.33% and 0.96% across seeds), with ours winning or tying in every seed. Visualizations are provided on our anonymous robot experiment page:

https://anonymous.4open.science/r/ICLR_Rebuttal-36184/01_rlbench/README.md

The learned local directions and graspable regions are shown under "Creating Task-equivalence Fiber", and diverse verified grasps with their successful executions under "Grasp Variations in Learned Vertical Direction", for both drawer and microwave. These results directly address the reviewer's concern by demonstrating that our framework can identify task-preserving equivalence from observed interactions without predefined symmetry groups or known equivalence-preserving transformations. Beyond this, as noted in the conclusion, implementation of tasks with more complex equivalence structures, including various other robotics tasks, remains an important direction for future work.

 


  • W2&W3 (Inconsistent notation and insufficient main-text clarity):

We thank the reviewer for highlighting the notation inconsistencies and insufficient methodological detail in the main text. We have revised the manuscript to explicitly connect the domain-specific implementations, including the MNIST encoder EcE_c, to the general quotient map qcq_c (Sec. 4.1), and standardized the corresponding notation and cross-references. To improve self-sufficiency, we have also clarified the local constant-rank assumption (Sec. 3.3), explained the implicit correction in Eq. (10) (Sec. 3.4), and incorporated the end-to-end training and inference procedure into the main text (Sec. 3.6).

 


References:

[1] James, Stephen, et al. "Rlbench: The robot learning benchmark & learning environment." IEEE Robotics and Automation Letters 5.2 (2020): 3019-3026.

AuthorsOpenReview ↗

Response to Reviewer muxE (2/3)

  • Q1 (Computational cost):

The SVD for the vertical/horizontal decomposition in Eq. (9) operates on the Jacobian in the relevant realization coordinates, not the full ambient data dimension. Specifically, the Jacobians are 64×564\times5 for MNIST and 6×76\times7 for Panda. For a Jacobian J∈Rm×nJ\in\mathbb{R}^{m\times n}, dense SVD has complexity O(mnmin⁡(m,n))O(mn\min(m,n)). In practice, the measured decomposition times are only 0.21 ms for MNIST and 0.42 ms for Panda, indicating modest computational overhead for the evaluated dimensions. This cost depends on the realization and quotient dimensions, not the ambient data dimension DD. The main additional costs are local chart construction and the implicit Newton correction at inference. In a separate timing measurement on four RTX 5090 GPUs, Panda flow training takes 7.0 min for ours vs. 24.1 min for ambient FM, and generating 2,048 samples takes 0.359 s vs. 3.603 s for FM with final projection, after a shared setup of about 124 s.

 


  • Q2 (Regarding correction term hih_i):

The vertical displacement EiVξE_i^V\xi in Eq. (10) preserves the quotient coordinate only to first order, and finite displacements can deviate from the curved fiber. The horizontal correction EiHhi(ξ)E_i^Hh_i(\xi) compensates for the fiber's nonlinear geometry, ensuring that finite displacements remain on the selected equivalence fiber. We agree that this explanation is important and have incorporated it into Section 3.4 of the main manuscript.

 


  • Q3 (Sensitivity):

We conducted a sensitivity analysis on the spatial-arm experiment, examining inaccurate corrections and the fraction of regular configurations retained. We evaluated the actual singular-value conditioning safeguard used by the implicit decoder, keeping the Newton tolerance, trained checkpoints, and sampling budgets unchanged.

We first swept ten thresholds across eight held-out requests and three seeds.

ThresholdSR (%) ↑SW₁ ↓Rank retained (%)Correction failures (%)Shortage (%)
10−1210^{-12}95.1800.0211100.0000.0000.000
10−1010^{-10}95.1800.0211100.0000.0000.000
10−810^{-8}95.1800.0211100.0000.0000.000
10−710^{-7} (default)95.1800.0211100.0000.0000.000
10−610^{-6}95.1800.0211100.0000.0000.000
10−510^{-5}95.1800.0211100.0000.0000.000
10−410^{-4}95.1800.0211100.0000.0000.000
10−310^{-3}95.1800.0211100.0000.0000.000
10−210^{-2}95.1800.0211100.0000.0000.000
10−110^{-1}87.8760.0497085.65315.0598.655

We further confirmed the default threshold against 10−410^{-4} using all 128 held-out requests across three seeds, with 2,048 requested outputs and the original 16,384-proposal limit.

ThresholdSR (%) ↑SW₁ ↓Rank retained (%)Correction failures (%)Shortage (%)
10−710^{-7} (default)95.1800.0211100.0000.0880.000
10−410^{-4}95.1800.021199.9980.0880.000

SR remained unchanged at 95.180%, with an SW₁ difference of only 5.35×10−75.35\times10^{-7} and no output shortages. Generated outputs were identical in 383/384 seed–request pairs. In contrast, an excessively restrictive threshold of 10−110^{-1} reduced SR to 87.876% and increased SW₁ to 0.04970, demonstrating robustness within the tested range but degradation at extreme thresholds.

 


  • Q4 (Loss function):

The objectives serve complementary purposes. Behavioral losses preserve task-relevant distinctions, Jacobian supervision identifies invariant directions, and transition losses enforce identity, inverse, composition, coordinate, and geometric transport consistency. The motivation for each term, including L_id, L_comp, and L_coord, is now summarized in Tab. 7. The existing component ablation (Tab. 8) evaluates their contributions. We additionally conducted a loss-weight sensitivity analysis, independently scaling individual coefficients from 0.25× to 2× their reference values. Results show generally stable AUROC, subspace F1, and rank accuracy near the reference weights, with some degradation at extreme multipliers. The coefficients are fixed across all runs (Eq. (C.4)), and Fig. 4 (App. C.1.2) shows stable performance around these values.

AuthorsOpenReview ↗

Response to Reviewer muxE (3/3)

  • Q5 (MNIST generation):

We extended our framework to MNIST image generation (Sec. 4.4.2, Fig. 3, App. C.5) using the original trained encoder, kept frozen while training flow matching in its learned fiber coordinates. As shown under "Within-fiber Generation" on our anonymous MNIST generation page: https://anonymous.4open.science/r/ICLR_Rebuttal-36184/02_mnist/README.md, the method generates diverse transformations of the same digit while preserving the selected equivalence class, including orientation or position tasks shown in Fig. 2a.

We additionally demonstrate across-fiber generation with the original encoder, also kept frozen ("Across-fiber Generation" on the same page). Vertical directions, initialized from the image transformations and projected onto the null space of the encoder Jacobian, preserve digit identity through the implicit decoder in Eq. (10), while a flow learned in quotient coordinates is lifted to image space through Eq. (9), enabling transitions between digits.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Revisions

We thank the reviewers and the AC for their careful evaluation. Below, we summarize the main concerns and how we addressed them.

  • Beyond explicit group actions and established benchmarks (muxE, WHj2). For this rebuttal, we ran robotics manipulation experiments of real task on RLBench opendraweropen_drawer and openmicrowaveopen_microwave sequences, where equivalence is learned from simulated rollouts without predefined transformations. From simulated grasp rollouts alone, the learned behavior map defines task-equivalence fibers with their vertical and horizontal directions. Ours improves opening success over standard FM, and the robot experiment page visualizes the learned fibers and the resulting grasps: https://anonymous.4open.science/r/ICLR_Rebuttal-36184/01_rlbench/README.md
  • MNIST generation (muxE). With the original encoder frozen, we generate different images within a fiber and transitions between digits across fibers, with visualizations in the experiment page: https://anonymous.4open.science/r/ICLR_Rebuttal-36184/02_mnist/README.md
  • Clarity and presentation (muxE, 3G6G). We made the main text more self-contained and consistent by stating key assumptions explicitly, unifying notation, and bringing the training and inference procedure into the main text.
  • Robustness and cost (muxE, 3G6G). We added analyses showing that the learned geometry is stable, the results are robust to numerical and loss settings, and the computational overhead is modest.
  • Generality and evaluation (WHj2, 3G6G). We clarified that our framework only requires the outcome to preserve, not the transformations that preserve it, and that success rate and distribution accuracy are complementary rather than conflicting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.