VineLM: Fine-Grained Model Selection for Agentic Workflows

A new workflow manager called VineLM enables fine-grained control over LLM stages in agentic workflows. Unlike existing managers that assign a static model per workflow, VineLM selects models per stage invocation based on runtime objectives like accuracy, cost, or latency. It uses a trie of model-choice prefixes and checkpointing to estimate performance without exhaustive profiling. At runtime, it re-roots the trie after each stage and replans dynamically. The paper is available on arXiv.