What they did

The authors designed a similarity-based mechanism that segments conversation history into windows for constructing the prompt used at each turn. This is intended to retain relevant information while limiting confusion from unrelated prior dialogue.

They paired this mechanism with two evaluation metrics that separate context-construction accuracy from the router’s ability to select a model. SWRouter was compared with individual language models and existing multi-turn routing baselines on dialogue benchmarks.

Key findings

• SWRouter achieved 16.26% higher evaluation accuracy than the best individual large language model, as reported by the authors.
• It produced an additional 8.22% gain over the Conv-ID Context baseline.
• The experiments indicate that context segmentation materially affects routing results in multi-turn dialogue.
• Separating prompt-construction quality from model-selection quality allowed the authors to assess where routing errors originated.

Why it matters

The results suggest that multi-turn model routing is not simply single-turn routing with the full chat history attached. Routing systems need to manage conversational context explicitly and evaluate that process independently, which is relevant to agent harnesses that choose among models with different capabilities or costs.

Caveats

The supplied abstract does not identify the evaluated model pool, benchmark names, computational costs, latency, or whether the reported percentage gains are relative or absolute. The evidence is therefore limited to the tested dialogue settings, and it remains unclear how the approach scales to very long conversations or transfers to different model collections and domains.