SAFformer:Improving Spiking Transformer via Active Predictive Filtering

A new spiking neural network architecture called SAFformer has set performance records on multiple image classification benchmarks while operating with sharply reduced energy consumption, according to a paper posted on the arXiv preprint server [1]. The model, detailed in a submission last revised on June 12, 2026, is built around what its authors call an active predictive filtering paradigm [1]. Rather than passively processing all incoming visual data, SAFformer is designed to suppress predictable signals and concentrate computation on salient features, an approach inspired by the brain's predictive coding mechanism [1][2]. The work was led by Zequan Xie [1]. On the CIFAR-10, CIFAR-100, and CIFAR10-DVS datasets, the architecture established new state-of-the-art results [1][2]. On the larger ImageNet-1K benchmark, SAFformer reached 80.44% Top-1 accuracy using 26.58 million parameters and consuming 5.88 millijoules of energy [1][2]. The authors describe this as an exceptional balance between accuracy and efficiency [2]. Spiking Neural Networks, the broader family to which SAFformer belongs, have drawn research interest because their event-driven computation mimics biological neurons and can yield substantial energy savings compared to conventional artificial neural networks [2]. The paper argues that existing Spiking Transformers have been limited by a passive reactive paradigm that struggles to isolate task-relevant information and wastes computation on redundant visual data [1][2]. The preprint appeared on arXiv, an open-access repository that hosts electronic preprints across physics, mathematics, computer science, and related fields [6]. Founded in 1991, the repository now receives roughly 24,000 submissions per month and has surpassed two million total articles [6]. Papers on arXiv are moderated but not peer-reviewed before posting [6]. The SAFformer manuscript was submitted on May 8, 2026, and updated a month later [1]. The paper's abstract page also links to several community-developed tools through the arXivLabs framework, including the Bibliographic Explorer for navigating citation trees and the CORE Recommender for surfacing related open-access research [4][5]. arXivLabs, formalized in 2020, allows third-party collaborators to build experimental features on top of the repository while adhering to arXiv's values of openness, community, excellence, and user data privacy [4].