HyScale is a long-context training system for learned sparse attention, the architecture family behind million-token contexts that has been in production since DeepSeek-V3.2. The paper's argument is that selection, not the attention math, is the bottleneck: for every query the system scores every previous token even though sparse attention reads only k, and current stacks materialize a full score table and then repeat the selection during recomputation.
The fix is to treat the output of selection as a set of indices rather than an activation. HyTopK finds each query's exact top-k on the fly without a score table, and selection replay reuses the forward pass's selection during recomputation. With those two changes plus end-to-end system work, GLM-5.1 training on 1M-token sequences runs 2.07x faster than Megatron-LM on 512 H20 GPUs. The same changes speed up the CSA layers of DeepSeek-V4-Flash by 2.01x and the QSA layers of Qwen3.8-Flash-Next by 1.57x at 1M tokens on 32 H800 GPUs.
The limitation is scope: these are training-throughput numbers on named hardware and model configurations, and the abstract reports no quality measurement, so any trade-off in model quality is not stated. The gains also concentrate where a stack was already materializing score tables, which is precisely the pattern HyTopK removes.
https://arxiv.org/abs/2610.32057
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.