Evaluate KGFT on broader downstream tasks

Investigate whether Kernel-Guided Feature Transform (KGFT) provides consistent benefits beyond image classification when applied to object detection, semantic segmentation, and parameter-efficient fine-tuning tasks.

Background

Kernel-Guided Feature Transform (KGFT) is evaluated in the paper primarily on image classification with ResNet and ViT architectures and on arithmetic reasoning with LLaMA-7B. The authors state that further evaluation is needed on additional downstream tasks, specifically object detection, semantic segmentation, and parameter-efficient fine-tuning, to determine whether the observed benefits of kernel-guided geometry generalize beyond the settings studied.

References

In future work, we will further explore the application of KGFT to a broader range of downstream tasks, such as object detection, semantic segmentation, and parameter-efficient fine-tuning, to investigate whether kernel-guided geometry can provide consistent benefits beyond classification.

— Dual-Manifold Geometry Guided Representation Learning: Adaptive Coupling between Kernel and Data Spaces  (2608.12737 - Zhang et al., 13 Aug 2026) in Conclusion, Future Work paragraph

Detection and segmentation instead need features at many spatial resolutions, so MCL could instead be added at multiple stages of the backbone so it can generate kernels at different resolutions. We have not tested this multi-stage setting and leave it for future work.

— MCL: Meta Convolution Layer  (2610.11117 - Reza et al., 8 Oct 2026) in Section Conclusion