architecture · first seen 31 Oct 2026
TopK sparse autoencoder
Also: TopK, TopK SAE
A sparse autoencoder that keeps only the k largest latent preactivations per input.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Sparse Autoencoder Attribution Graphs Are Decided at the MarginPapers · 31 Oct 2026 · Karl Li
- How Small Can We Go? Calibrating Activation-Cache Compression for SAE TrainingPapers · 5 Nov 2026 · Abraham Yeung
- Cross-Layer TopK: Dynamically Allocating Sparsity Across Layers in TranscodersPapers · 6 Nov 2026 · Anyuan Zhuo, Shaohua Zhang, Tianle Li, Yuan Lin, Diandian Gu, Jun Wang, Pinyan L…
Latest sources
- Cross-Layer TopK: Dynamically Allocating Sparsity Across Layers in TranscodersPapers · 6 Nov 2026 · Anyuan Zhuo, Shaohua Zhang, Tianle Li, Yuan Lin, Diandian Gu, Jun Wang, Pinyan L…
- How Small Can We Go? Calibrating Activation-Cache Compression for SAE TrainingPapers · 5 Nov 2026 · Abraham Yeung
- Sparse Autoencoder Attribution Graphs Are Decided at the MarginPapers · 31 Oct 2026 · Karl Li