CoLLM: A Unified Framework for Co-execution of LLMs Federated Fine-tuning and Inference

This feed contains multiple 2026 arXiv papers on systems and ML infrastructure: CoLLM — a unified co-execution framework that merges federated PEFT fine-tuning and inference on shared edge replicas to improve goodput; MSGAF — multi-scale graph fusion for scene-aware microservice latency estimation; Spot-and-Scoot — a low-cost probing technique that infers spot-instance availability by submitting-and-cancelling provisioning requests; B-PASTE — beam-aware speculative execution for LLM agents that speculates bounded future branches under resource constraints; KAIROS — context-aware power and GPU/