Computer ScienceDeepSeek Cuts KV Cache Footprint FourfoldA causal encoder-decoder MoE pairs cross-layer sparse-attention reuse with FP4 caching to reduce global context memory to 890 bytes per token.cs.CLSep 19, 20264 min2609.19969
PhysicsXENONnT Extends Light Dark Matter Limits With S2-Only DataA 7.83 tonne-year ionization-only analysis models low-energy backgrounds in four observables and excludes spin-independent scattering above 6.0×10⁻⁴⁵ cm² at 5 GeV/c².astro-phhep-exhep-phAug 3, 20264 min2601.11296
