Routing between language models is usually studied for single-turn queries. This paper looks at multi-turn tasks, where the best model may change as the conversation evolves. The authors introduce 'handoff curves' as a tool to record task success and cost across different routing decisions.
The curves are used to compare several router architectures. The study focuses on three decisions: which model to start with, when to switch to another model, and how supervision targets affect training. By plotting success against cost, the curves aim to reveal trade-offs that simpler metrics miss.
Because the abstract is truncated, the specific findings are not available here. The contribution appears to be the handoff-curve methodology itself, which could help future work on multi-turn LLM routing.