Multi-turn LLM routing requires careful handling of dialogue history to avoid information loss or confusion. SWRouter addresses this with a similarity-based context segmentation that constructs prompts by retaining only relevant prior turns, and a dual-metric evaluation that separates prompt quality from model selection. On multi-turn dialogue benchmarks, SWRouter achieves a 16.26% improvement in evaluation accuracy over the best single LLM and an additional 8.22% gain over the Conv-ID Context baseline. This is interesting because it demonstrates that effective multi-turn routing demands joint design of context construction and evaluation, rather than simply extending single-turn approaches.
Similarity-aware context windows improve model routing across multi-turn conversations
SWRouter jointly segments dialogue history and selects among language models, while evaluating prompt construction separately from routing quality.
Chinese Tech
Yu Wang · Yuchen Li · Rui Kong · Xinran Chen · Jiamin Chen · Hengyi Cai · +8 more
Shanghai Jiao Tong University · Baidu Inc.
Research Digest··2 min read
In one line
SWRouter improves multi-turn LLM routing accuracy by 16.26% over the best individual LLM.
Before you read this
Two ideas this paper builds onRead it as
Wang et al. introduce SWRouter, a method for routing each turn of a multi-turn conversation to a suitable large language model.
The authors designed a similarity-based mechanism that segments conversation history into windows for constructing the prompt used at each turn.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.