SWRouter: Similarity-Contractive Window Routing for Multi-Turn Large Language Model Conversations

2609.11414 ↗ · 2026-09-10 · cs.CL, cs.AI, cs.IR

Yu Wang, Yuchen Li, Rui Kong, Xinran Chen, Jiamin Chen, Hengyi Cai, Shuaiqiang Wang, Jiashu Zhao, Yulun Zhang, Zhonghao Lyu, Haoyi Xiong, Linghe Kong +2

The finding

Headline claim

Routing multi-turn LLM conversations with similarity-based context windows plus a contrastively trained router beats both the best single model and naive context handling, and prompt-construction quality must be measured separately from model-selection quality.

The test

Multi-turn dialogue benchmarks with a pool of candidate LLMs (GPT, DeepSeek, Claude, LLaMA, Gemini families referenced); SWRouter vs routing baselines (RouterDC-style), a Conv-ID Context baseline, and the best individual LLM, with in-distribution and OOD splits and a decoupled three-metric evaluation (evaluation accuracy, construction accuracy, router performance).

Effect size & scope

16.26% average improvement in evaluation accuracy over the best individual LLM, 8.22% over the Conv-ID Context baseline in-distribution, and 2.68% over Conv-ID Context on OOD tasks; gains attributed by the authors' decoupled metrics mainly to improved construction accuracy, with router performance above 1 across settings. Exact dataset sizes and judge-model identity are not visible in the provided text.

Limitations

Quality scores come from a judge model rather than human raters, so measured gains inherit judge bias; candidate model pool, benchmark details, and cost/latency tradeoffs of running multiple candidate models are not fully specified in the provided text; gains are on benchmark dialogues, not production traffic.

Counterpoint

A judge-LLM-scored benchmark with a hand-built Conv-ID baseline invites benchmark gaming and prompt sensitivity; picking the best of several frontier models per query multiplies inference cost, so end-to-end value depends on overhead numbers not shown; the routing gain partly restates that better prompts help all models.

Economy implication

Little or nothing yet for the aggregate AI economy; at most it signals that if routing matures, it strengthens multi-model competition at the API layer by making per-query model switching practical, which bears on where margin settles between labs and middleware, a claim that needs cost data to test.

Who this matters for

Technical Capability: Shows that multi-turn routing requires joint context-construction and evaluation design, so single-turn routing methods do not transfer unchanged.

Market & Valuation: If routing lowers switching costs between frontier models per query, it erodes any single lab's pricing power in conversational API markets.

Bears on (advisory · papers never write evidence)

[5.3] API switching costs are low (customers multi-home and swap easily).

Abstract

Large language models exhibit complementary strengths, motivating routing methods that dispatch each query to the most suitable model. Although existing routers are effective in single-turn settings, they do not directly transfer to multi-turn dialogue, where routing performance critically depends on how historical context is segmented, retained, and incorporated into the current prompt. This introduces two fundamental challenges: preventing information loss and information confusion during context construction, and evaluating routing quality without conflating model selection with prompt construction quality. In this paper, we propose SWRouter, a Similarity-Contractive Window Router for multi-turn large language model routing. SWRouter combines a similarity-based context segmentation mechanism for prompt construction with a dual-metric evaluation framework that decouples construction accuracy from router performance. Experiments on multi-turn dialogue benchmarks demonstrate that SWRouter consistently surpasses strong baselines, achieving a 16.26% improvement in evaluation accuracy over the best individual large language model and an additional 8.22% gain over the Conv-ID Context baseline. Our results highlight that multi-turn large language model routing requires a joint design of context construction and evaluation, rather than a direct extension of single-turn routing methods.