Nexus: Depth-Adaptive KV-Cache Splicing and Retrieval-Decoupled Tool Routing for Agentic LLMs on Unified Memory

arXiv:2608.20397v1 Announce Type: new Abstract: Agentic large language models LLMs on the Model Context Protocol MCP reencode verbose tool schemas every turn, so prefill quadratic in sequence length dominates timetofirsttoken TTFT as the tool registry grows.