The most significant papers right now, ranked by engagement.
A 260-hour benchmark spanning 21 attacks and five emotions shows conventional detectors can approach chance-level performance under emotional spoofing.
A cone-preservation proof handles one fast and two identical slow servers, closing the first nontrivial case beyond Lin-Kumar’s two-server model.
A compact knowledge-base backbone queries full user histories at each layer, improving offline ranking metrics while cutting per-example FLOPs by 8×–25×.
Module-specific NVFP4 quantization covers projections, optimizers, and attention, reaching a 1.47% loss gap on 3B/64B-token pretraining.
