An arXiv preprint reports that Core-KAN, an adaptive operator for computer vision, posted higher scores than ResNet-50 in tests on ImageNet-1K, COCO and ADE20K, with additional reported margins over competing methods in the comparison tables.
Separating scale from content
Core-KAN treats spatial adaptation as two linked but separate decisions: the geometric scale and structure of a filter, and the way local image content mixes different filtering patterns. Its scale controller chooses the relative scale used to query a shared continuous kernel field, while an independent mixing controller combines bases according to local content.
The operator predicts a dense scale field, normalizes it against an exponential moving average reference, evaluates a shared continuous kernel field, samples fixed-size kernels and interpolates neighboring support responses. This creates a response bank—stored responses at reference scales—that can be read by interpolation rather than by synthesizing a separate kernel at every position.
The default setup uses a 3-by-3 spatial support, 16 basis fields and eight scale supports—reference points at which the field is sampled—with scale bounds from 0.5 to 2.0. The authors describe eight supports as a practical trade-off; using 16 improves interpolation fidelity but increases response-bank cost.
Reported gains across three benchmarks
On ImageNet-1K classification, Core-KAN reported 81.45% top-1 accuracy, meaning its first-choice label was correct at that rate, and 95.68% top-5 accuracy, meaning the correct label appeared among its five leading choices. It used 26.61 million parameters. The paper reports relative gains of 3.84% and 1.53% over ResNet-50, respectively, alongside a 4.11% parameter overhead; compared with KernelWarehouse, it reports a 0.49% relative gain with 73.92% fewer parameters.
On COCO, using Mask R-CNN with FPN for object detection and instance segmentation, Core-KAN reported 43.0 box AP and 39.5 mask AP under the 1×, 12-epoch schedule. Under the 3×, 36-epoch schedule, it reported 46.2 box AP and 42.2 mask AP. The paper reports relative improvements over ResNet-50 of 13.46% and 14.49% in the first comparison, and 12.96% and 13.75% in the longer one.
In the longer-schedule COCO table, Core-KAN ranked first on five of six reported box metrics and all six mask metrics. It did not lead every entry: the paper reports a 0.17% gap on the large-object box metric, listing 58.9 for Core-KAN and 59.0 for KernelWarehouse.
On ADE20K semantic segmentation, Core-KAN reported 44.19% mean intersection-over-union, a measure of overlap between predicted and reference regions, and 54.38% mean accuracy. It used 68.50 million parameters; the paper reports relative gains over ResNet-50 of 10.47% for mIoU and 9.61% for mAcc, margins over FDConv of 1.59% and 1.40%, 3.79% parameter overhead, and 2.14% fewer parameters than FDConv.
The comparisons have limits
The authors also examined how the learned kernels changed with scale. On a fixed 3-by-3 sampling grid, normalized weights showed scale-dependent changes, while a principal-component analysis—a way of reducing many measurements to a few summary directions—traced ordered, basis-specific trajectories. The authors interpret these patterns as continuous and interpretable kernel transformations.
The work is a preprint on arXiv, identified as arXiv:2608.19817v1 and dated 20 August 2026. The authors describe the ImageNet table as a reference comparison because the reported or reproduced training recipes differ across methods, so those percentages are not a clean like-for-like test.
The authors also flag a practical constraint: the response bank becomes more expensive as the number of scale supports grows. Their open question is whether that cost can be reduced without weakening the interpolation fidelity that improved when support count rose from eight to 16.
Paper data and sources
Original title: Core-KAN: Continuous Vision Kernels with Kolmogorov-Arnold Networks
Authors: Lan Guo, Mengling Li, Haoran Li et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text