AWhat changed
Reformulated recommendation around sequentially unified user actions and introduced HSTU blocks with pointwise aggregated attention, multiplicative gating, ragged fused kernels and stochastic sequence-length training.
Scientific scope
Reformulated recommendation around sequentially unified user actions and introduced HSTU blocks with pointwise aggregated attention, multiplicative gating, ragged fused kernels and stochastic sequence-length training. Attention uses SiLU/normalization rather than softmax; feature and streaming assumptions matter. Stochastic-length compute savings alter sampled histories and have empirical quality checks, not universal accuracy preservation. Power-law scaling and online A/B metrics are source-reported on selected systems; comparisons with LLM compute do not establish identical tasks or training regimes.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Attention uses SiLU/normalization rather than softmax; feature and streaming assumptions matter. Stochastic-length compute savings alter sampled histories and have empirical quality checks, not universal accuracy preservation. Power-law scaling and online A/B metrics are source-reported on selected systems; comparisons with LLM compute do not establish identical tasks or training regimes.
Contributors
- Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.
Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction