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
PhysicsPressure Reveals a Fragile Bright Magnetic ExcitonHydrostatic compression extinguishes NiPS3 photoluminescence by 1.5 GPa without magnetic, structural, or electronic reconstruction.quant-phcond-matphysicsAug 11, 20264 min2607.27695
