Demote the repo-exploration step to a cheap model and keep 78-94% of localization quality for 5-16% of the tokens

arXiv 2608.29675·high signalXBlueskyLinkedInCopy link

IssueLoc-Bench evaluated five explorer models under an identical read-only interface on 499 SWE-bench Verified tasks plus 500 tasks from 153 other repositories, measuring the file-finding stage separately from patching. Lower-cost explorers retained roughly 78-94% of the reference Hit@3 and 73-92% of its F1 while cutting mean agent time 41-88% and token usage 84-95%. The operating point you want depends on the downstream handoff: ranking metrics matter if the next stage re-reads candidates, F1 and exact match matter if localization is a hard file gate.