acceptodds
Under review as a conference paper at ICLR 2027

DGKAN:Dynamic-Grid Kolmogorov-Arnold Networks

Abstract

We introduce DGKAN, a dynamic-grid Kolmogorov-Arnold Network that learns to redistribute a fixed number of grid points across the input domain. Existing KAN use uniformly spaced, predefined grids, implicitly assigning the same representational resolution to all regions. This allocation can be inefficient when function complexity and feature distributions are heterogeneous, motivating the problem of how to optimally allocate a limited grid budget across the domain. DGKAN addresses this problem through backward-guided grid adaptation: output-layer grids are relocated according to local prediction errors, while intermediate-layer grids adapt to the empirical distribution of latent features. To support non-uniform grids, we develop adaptive radial basis functions with grid-dependent bandwidths and selectively introduce variational Bayesian modeling into the Radial Basis Function (RBF) branch for localized shrinkage and uncertainty estimation. Experiments and ablation studies demonstrate the effectiveness of dynamic grid allocation, while generalization tests across diverse tasks validate the generality of DGKAN.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Kolmogorov–Arnold Networks, Dynamic Grid, Adaptive Grid Allocation, learnable activation functions

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_cK1EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose DGKAN, which dynamically reallocates fixed grid point budgets to high complexity regions through a layerwise adaptation scheme that ensures stable relocation and function preservation.

Critical strengths

By formulating grid adaptation as a budget-constrained allocation problem, DGKAN dynamically reallocates fixed grid point budgets to high complexity regions, improving representational efficiency while preserving network composition.

Critical weaknesses

Contribution 2 represents the main methodological innovation proposed by the authors; however, several key design elements within it have not been individually validated.

There is a gap in the experimental design regarding "layerwise adaptation": the authors do not justify why hidden layers must be guided by the empirical density, while the output layer must be guided by the prediction residual.

The authors constrain the magnitude of each individual movement of grid points using a "relocation rate" and a "maximum displacement". This is a reasonable design choice for numerical stabilization. However, the paper does not discuss how training stability varies under different displacement limits.

Questions for authors

Could the authors explain why empirical feature distributions are used to guide hidden-layer adaptation while prediction residuals are used for output-layer adaptation, and provide evidence that this design is more effective than alternative configurations?

1 comment
AuthorsOpenReview ↗

Rationale and Stability Analysis of Layerwise Dynamic Grid Adaptation

1. Why use density in hidden layers and residuals at the output?

Response. We thank the reviewer for raising this point. Our asymmetric layerwise adaptation strategy is motivated by the different supervision signals available at the output and hidden layers.

At the output layer, the input representation is directly connected to the final prediction. Therefore, we can compute interval-wise prediction residuals and use them to construct a local error monitor for grid distribution.

Backpropagating the final prediction error to a hidden layer yields an error signal, not the true approximation residual of the hidden univariate function. It depends on downstream parameters, nonlinearities, and channel-wise scaling, and may therefore mix credit assignment with the local resolution requirement of the hidden basis.

We consequently use the empirical density of hidden features as an observable and lower-variance proxy for where representation resolution is needed. This density is indirectly shaped by the final prediction loss because the hidden representations are learned under that loss. Thus, the output-layer residual provides a direct task-aware signal, while hidden-layer density provides a stable representation-resolution signal. Density alone is insufficient to characterize prediction error, which is why we retain residual-driven adaptation at the output layer.

Due to the limited time available, we conducted a preliminary experiment on an NVIDIA RTX 3090 using a subset of the data (fpacketf_{packet}) and the original experimental settings. For each configuration, we used five different random seeds and repeated the experiment five times. The table reports the mean MSE and standard deviation across these five runs.

Hidden-layer/ outputLocal-uniform G=5G=5Local-uniform G=9G=9Local-conflict G=5G=5Local-conflict G=9G=9
Fixed / fixed0.014294 ± 0.0190000.006764 ± 0.0094770.030896 ± 0.0077770.004589 ± 0.008186
Density / density0.010341 ± 0.0136050.001896 ± 0.0015990.012578 ± 0.0054980.002594 ± 0.001413
Residual / density0.000700 ± 0.0001710.000240 ± 0.0002150.001450 ± 0.0005500.000420 ± 0.000180
Residual / residual0.002274 ± 0.0010020.000640 ± 0.0002680.001800 ± 0.0007590.000700 ± 0.000346
Density / residual, proposed0.000550 ± 0.0002000.000160 ± 0.0000600.001200 ± 0.0004500.000300 ± 0.000120

2. Relocation rate, maximum displacement, and stability

Response. We thank the reviewer for pointing out this omission. We use the relocation rate, denoted by λ\lambda, and the maximum per-update displacement, denoted by δmax⁡\delta_{\max}, to limit the magnitude of each grid update. These two parameters affect the trajectory of grid movement and should be distinguished from the final predictive performance.

We conducted controlled layerwise sensitivity experiments. With λ=0.2\lambda=0.2 fixed, changing δmax⁡\delta_{\max} from 0.250.25 to 2.02.0 resulted in relatively stable selected frozen-stage test errors across the 20 paired configurations, without an evident systematic trend. This suggests that, under the current experimental settings, the maximum displacement is not the primary factor determining the selected final error. With δmax⁡=0.25\delta_{\max}=0.25, changing λ\lambda from 0.20.2 to 1.01.0 yielded 9 paired improvements and 11 degradations, and increased the mean error in the local-uniform G=5G=5 condition. Thus, our current evidence does not support a universal optimum for either parameter.

We recorded the minimum grid spacing, number of grid updates, cumulative grid movement, immediate change in training-anchor loss, and checkpoint-level performance for every run. For the proposed density/residual guidance, we observed 720 grid-update events in total. Only 174 events (24.2%) immediately reduced the training-anchor error, whereas 546 events (75.8%) temporarily increased it. Nevertheless, after the subsequent frozen-grid continuation, all runs achieved lower final test MSE than their corresponding pre-update states. This pattern supports our interpretation that a grid relocation should not be judged solely by its immediate loss change. Because coefficient projection is approximate, an individual relocation can temporarily perturb the local representation, while subsequent retraining can compensate for this perturbation and yield a better final solution.

In the revised manuscript, we will add the complete sensitivity results and grid-movement trajectories. Our results support a conditional-effect interpretation: the relocation constraints are useful for controlling representation perturbations, but their predictive impact depends on the data distribution, grid budget, and update schedule.

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

This article introduces a dynamic KAN network. The author pays attention to the inefficiency caused by fixed nodes in KAN and proposes a dynamic optimization method for KAN.

Critical strengths
  1. This article is written clearly, and dynamically allocating network nodes under a fixed grid budget seems reasonable.
  2. The combination of output residue guided+hidden layer density guided adaptation has certain innovative methods, and is matched with ARBF and coefficient transplantation. Technically, it is not simply a modification of KAN. 3.The experiment is complete and can be followed. The ablation experiment is also convincing.
Critical weaknesses
  1. I admit that a renewed call for community attention to KAN requires such an article, but technically, its novelty is not sufficient to support publication in ICLR. This article actually introduces adaptive mesh/non-uniform grid allocation into KAN, lacking a strong comparison with more direct learnable knot and adaptive basis/grid optimization methods, making it difficult to prove that this is a new learning principle for the machine learning community.
  2. Although the experiment is complete, the existing experimental strength does not support generality. The existing experimental results mainly demonstrate the use of artificially constructed 1D function fitting; Darcy only presented limited results, and the improvement in time series prediction compared to TimeKAN was small and inconsistent; for example, the average results of ETTh1/ETTm1 were basically the same or even partially worse.
  3. The Bayesian module lacks empirical closed loops. The paper emphasizes epidemiological uncertainty, but does not show key ablations of calibration, NLL/ECE/coverage, or the Bayesian branch, so this part is more like an add-on component rather than a validated contribution.
Questions for authors

Major comments can be followed in the weakness.

Minor comments:

  1. The grid adaptation procedure introduces several additional hyperparameters, including the relocation rate, maximum displacement, minimum spacing, and update frequency. Please provide a clearer sensitivity analysis or at least discuss how these values are selected and whether performance is sensitive to them.

  2. The computational overhead of dynamic grid relocation and coefficient re-projection is not fully characterized on the downstream tasks. Reporting training time, peak memory, and inference cost relative to the corresponding KAN baselines would make the efficiency claim more convincing.

  3. Please clarify how often grid updates are performed during training and whether the final performance depends on the update schedule. It would also be useful to report whether the grid locations converge or continue to fluctuate near the end of training.

1 comment
AuthorsOpenReview ↗

Response to Reviewer CktN: Novelty, Generality, Uncertainty, Stability, and Computational Cost

1. Novelty relative to learnable grids and adaptive bases

Response. We thank the reviewer for raising this concern. Our question is narrower: under a fixed grid budget, can dynamic allocation outperform a static uniform grid?

DGKAN performs grid relocation as a periodic non-gradient operation. Candidate locations are obtained from hidden-feature density or output residuals, followed by bounded movement, minimum-spacing constraints, and coefficient re-projection. Because grid coordinates are not differentiated through the full network, this avoids storing additional gradient graphs for grid optimization and reduces the associated memory overhead.

To our knowledge, DGKAN is among the early studies of fixed-budget, layerwise dynamic grid allocation in KAN. We will add relevant concurrent work and comparisons when suitable implementations are available.

2. Limited generality and inconsistent time-series performance

Response. We added ETT experiments with standard KAN, RBF-KAN, hidden-layer-only DGKAN, and output-layer-only DGKAN. All runs used the same preprocessing, training, and evaluation protocols on an NVIDIA RTX 3090. Each setting was repeated with five random seeds; the table reports mean MSE.

DatasetHorizonKANRBFKANDGKANhiddenDGKAN_{hidden}DGKANoutputDGKAN_{output}
ETTh1960.3745890.3756080.3754930.375456
ETTh11920.4199970.4235510.4250390.424825
ETTh13360.4756900.4718160.4727010.472672
ETTh17200.4686320.4709320.4712510.471138
ETTm1960.3318980.3252050.3256850.325672
ETTm11920.3642220.3630600.3633180.363559
ETTm13360.3952770.3938320.3938010.393645
ETTm17200.4508130.4586650.4582140.457975
ETTh2960.2939320.2922240.2922550.292271
ETTh21920.3772450.3738680.3740800.372162
ETTh23360.4224320.4208150.4208150.420815
ETTh27200.4358190.4344220.4344220.434422
ETTm2960.1769740.1764080.1764050.176407
ETTm21920.2435500.2431070.2431020.243108
ETTm23360.3056700.3035480.3035330.303535
ETTm27200.4023110.4008360.4007680.400816

These variants separate RBF parameterization from layerwise adaptation. Results are not uniformly better because DGKAN is inserted without changing the surrounding architecture, data pipeline, optimizer, or training schedule. This controlled protocol measures transferability, not the maximum after task-specific tuning. We provide additional Darcy configurations in the supplement and plan to add more operator-learning scenarios and datasets.

3. Bayesian component, uncertainty evaluation and epidemiological

Response. The Bayesian variational RBF branch is optional. Its purpose is to reduce the effect of isolated outliers and abnormal local coefficient updates through prior-induced shrinkage, while the deterministic backbone models global trends.

The noise experiment suggests stable prediction as noise increases, but does not establish calibrated uncertainty or epidemiological validity. The manuscript does not emphasize epidemiological uncertainty or use epidemiological modeling as a central claim. If suitable data and protocols become available, we may add a real epidemiological experiment in future work.

We will add a Bayesian-off ablation and, where applicable, report NLL, CRPS, coverage, and calibration error, while separating variational shrinkage from grid adaptation and reporting any lack of consistent Bayesian advantage.

4. Hyperparameter sensitivity and Update frequency

Response. We use relocation rate λ\lambda and maximum displacement δmax⁡\delta_{\max} to control grid movement. With λ=0.2\lambda=0.2, changing δmax⁡\delta_{\max} from 0.250.25 to 2.02.0 produced stable selected frozen-stage errors across 20 paired configurations, without a systematic trend. With δmax⁡=0.25\delta_{\max}=0.25, changing λ\lambda from 0.20.2 to 1.01.0 yielded 9 improvements and 11 degradations and increased the mean error for local-uniform G=5G=5; no universal optimum is supported.

For the proposed guidance, 720 updates were recorded: 174 (24.2%) immediately reduced training-anchor error and 546 (75.8%) temporarily increased it. Nevertheless, all runs achieved lower final test MSE after frozen-grid continuation. Thus, immediate loss changes and final performance should be evaluated separately because approximate coefficient projection may be compensated by subsequent retraining. We will add complete sensitivity results and grid trajectories.

5. Computational overhead

Response. Under the same time-series protocol, total training time increased from 140.10 s for RBF to 143.12 s for DGKAN, an increase of 3.02 s or approximately 2.16%. The additional cost was limited in the tested setting.

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

This paper introduces DGKAN, a dynamic-grid Kolmogorov-Arnold Network that learns to redistribute a fixed budget of grid points across the input domain, addressing the inefficiency of uniformly spaced predefined grids that assign equal representational resolution to all regions regardless of heterogeneous function complexity and feature distributions. Experiments and ablation studies demonstrate the effectiveness of dynamic grid allocation. The paper is well organized structurally. The main issue is that the experiments are insufficient, lacking experiments on real data, complexity analysis, ablation studies, and other evidence to demonstrate the effectiveness of each proposed module.

Critical strengths
  • The paper is well structured and easy to follow, and the motivation is clearly articulated.
  • Inspired by adaptive mesh refinement methods, the paper addresses the problem that data are often unevenly distributed across grid intervals, with densely sampled regions providing richer supervisory signals that enable the model to learn more detailed representations in those regions.
  • The paper proposes a dynamic grid method that adaptively redistributes grid points during training through a layer-by-layer adaptive scheme.
Critical weaknesses
  • The experimental validation is somewhat limited, as it was only verified on simulated experiments, scientific computing and time series. The effectiveness of the method should be validated on more real-world tasks.
  • The improvement of the experimental results over the comparison methods is too marginal, especially on the time series forecasting problem.
  • The method adds a Grid Adaptation module, and a complexity analysis should be conducted to see whether the increase in complexity is worthwhile.
  • The paper should experiment with more basis functions and compare with more KAN methods to demonstrate the effectiveness of the proposed dynamic grid method.
  • The grid sensitivity analysis should be conducted on real data to demonstrate the effectiveness of grid adaptation.
  • Some modules lack ablation experiments, such as Bayesian variational RBF modeling.
Questions for authors

See Weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer JRen

1. Real-world validation is limited, with marginal gains over baselines.

Response. We thank the reviewer for raising this concern.

We added ETT experiments with standard KAN, RBF-KAN, hidden-layer-only DGKAN, and output-layer-only DGKAN. All runs used the same protocol on an NVIDIA RTX 3090, with five random seeds; the table reports mean MSE.

DatasetHorizonKANRBFKANDGKANhiddenDGKAN_{hidden}DGKANoutputDGKAN_{output}
ETTh1960.3745890.3756080.3754930.375456
ETTh11920.4199970.4235510.4250390.424825
ETTh13360.4756900.4718160.4727010.472672
ETTh17200.4686320.4709320.4712510.471138
ETTm1960.3318980.3252050.3256850.325672
ETTm11920.3642220.3630600.3633180.363559
ETTm13360.3952770.3938320.3938010.393645
ETTm17200.4508130.4586650.4582140.457975
ETTh2960.2939320.2922240.2922550.292271
ETTh21920.3772450.3738680.3740800.372162
ETTh23360.4224320.4208150.4208150.420815
ETTh27200.4358190.4344220.4344220.434422
ETTm2960.1769740.1764080.1764050.176407
ETTm21920.2435500.2431070.2431020.243108
ETTm23360.3056700.3035480.3035330.303535
ETTm27200.4023110.4008360.4007680.400816

These variants separate RBF parameterization from layerwise adaptation. Results are task- and horizon-dependent because DGKAN is inserted without changing the surrounding framework. This measures transferability rather than task-optimized performance. We will add more forecasting tasks, operator-learning scenarios, and applications.

2. The computational cost

Response. Grid relocation is a periodic non-gradient operation outside the backpropagation graph. Relocation, spacing constraints, and coefficient re-projection use detached grid statistics, so grid coordinates require no stored gradients or activation graphs and introduce no persistent additional training-memory overhead.

Under the same time-series protocol, total training time increased from 140.10 s for RBF to 143.12 s for DGKAN (+3.02 s, 2.16%). Thus, the added cost is computation for grid statistics and projection rather than memory for backpropagation. We will report time, memory, latency, throughput, and separate update/projection costs.

3. More basis functions

Response. We agree that comparisons with additional bases and direct KAN variants are important. DGKAN is, to our knowledge, an early study of fixed-budget, layerwise dynamic grid allocation in KAN. Our controlled RBF/B-spline comparisons do not establish universal superiority.

As this area develops, we will add concurrent citations and comparisons when suitable implementations and protocols are available. The contribution is fixed-budget, layerwise allocation, not a universally optimal basis.

4. Grid sensitivity on real data is missing

Response. We performed preliminary controlled sensitivity experiments on relocation rate λ\lambda and maximum displacement δmax⁡\delta_{\max}. With λ=0.2\lambda=0.2, changing δmax⁡\delta_{\max} from 0.250.25 to 2.02.0 produced stable selected frozen-stage errors across 20 paired configurations, with no systematic trend. With δmax⁡=0.25\delta_{\max}=0.25, changing λ\lambda from 0.20.2 to 1.01.0 yielded 9 improvements and 11 degradations and increased the mean error for local-uniform G=5G=5; no universal optimum is supported.

For the proposed guidance, 720 updates were recorded: 174 (24.2%) immediately reduced training-anchor error and 546 (75.8%) temporarily increased it. Nevertheless, all runs obtained lower validation-selected test MSE after frozen-grid continuation. Immediate loss and final performance should therefore be evaluated separately because subsequent retraining may compensate for projection-induced perturbations.

These preliminary results do not replace comprehensive real-data analysis. We will test grid budget, λ\lambda, δmax⁡\delta_{\max}, update interval, and minimum spacing on real data, reporting multi-seed statistics, trajectories, and overhead.

5. Ablation study of the Bayesian variational RBF component Response

Response. We thank the reviewer for raising this concern. The Bayesian variational RBF branch is optional. Its purpose is to reduce the effect of isolated outliers and abnormal local coefficient updates through prior-induced shrinkage, while the deterministic backbone models global trends.

We will add a Bayesian-off ablation and, where applicable, report NLL, CRPS, coverage, and calibration error, while separating variational shrinkage from grid adaptation and reporting any lack of consistent Bayesian advantage.

In subsequent versions, we will further extend the ablation study to cover additional Bayesian settings, uncertainty objectives, and alternative coefficient-migration strategies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.