CRAFT: Cost-aware Expert Replica Allocation with Fine-Grained Layerwise Estimations

This batch of arXiv submissions covers system, ML-serving, and distributed-compute research with several security-relevant implications. Key points: CRAFT (expert replication for MoE) shows over-replication can waste GPU memory and cause resource contention and throughput degradation—risking resource exhaustion or DoS in model-serving fleets. Spark-LLM-Eval introduces large-scale, distributed LLM evaluation with response caching (content-addressable cache) that improves cost but could expose cached model outputs or enable data leakage if caching/backing stores are misconfigured. FedRefine (f¶d