Search papers, labs, and topics across Lattice.
This paper introduces SWE-Router, a novel routing mechanism for multi-turn agentic software engineering tasks that optimally balances the use of cheap and expensive large language models (LLMs). By allowing a less capable model to explore a task for a few turns and analyze the resulting trajectory before deciding whether to escalate to a more powerful model, SWE-Router enhances cost efficiency without sacrificing performance. The authors demonstrate that this approach adheres to a Bayes-optimality theorem, showing significant improvements in routing decisions compared to existing methods that rely solely on task descriptions.
Routing decisions based on exploratory trajectories can significantly boost cost efficiency in software engineering tasks without losing the performance edge of stronger models.
Large language models (LLMs) embedded in multi-turn agentic harnesses are reshaping software engineering (SWE), but routing every task to a frontier model is wasteful when many issues admit cheap fixes. Existing LLM routers operate on the task description alone, which inherits an information-theoretic Bayes-error floor in agentic settings: a similar issue can hide either a localized typo or a multi-module refactor, and the prompt does not separate the two. We introduce SWE-Router, a value-based temporal approach that lets a cheap model run for a few exploratory turns and reads the resulting partial trajectory before deciding whether to continue cheaply or to escalate to an expensive model. We provide a Bayes-optimality theorem showing that conditioning on the partial trajectory never harms routing and is strictly better whenever exploration is informative. Across the LLM pairs of weak and strong models spanning the contemporary cost--capability frontier, we show that SWE-Router greatly improves the cost efficiency of SWE tasks, while maintaining the majority of the performances of the stronger model. We additionally release a multi-LLM trajectory dataset which allows reproduction of our trajectory-level routing.
