According to a paper published on arxiv.org, researchers have introduced SWE-Router, a routing system designed to improve cost efficiency in multi-turn AI software engineering tasks. The approach addresses a key limitation: routing decisions based solely on task descriptions cannot distinguish between simple fixes and complex refactors that may appear similar in the initial prompt.
SWE-Router uses a “value-based temporal approach” that allows a cheaper model to run for several exploratory turns before deciding whether to continue with the cheap model or escalate to an expensive one, according to the paper. The researchers provide a “Bayes-optimality theorem showing that conditioning on the partial trajectory never harms routing and is strictly better whenever exploration is informative.”
According to arxiv.org, across multiple LLM pairs spanning the “contemporary cost-capability frontier,” SWE-Router “greatly improves the cost efficiency of SWE tasks, while maintaining the majority of the performances of the stronger model.” The paper is scheduled to appear at the 5th Deep Learning for Code Workshop at ICML 2026.
The research also released a multi-LLM trajectory dataset to enable reproduction of trajectory-level routing experiments. This work contributes to a broader trend in AI routing systems, including ComplianceGate, which arxiv.org reports achieved “39% median latency reduction” and “33-52% cost savings” through classifier-gated routing in regulated industries.