7B Repo Scout Matches Top AI Fixer at One-Fifth Cost

SuperScout, a 7B repo-scouting model, solved 159 of 266 SWE-bench Pro tasks. The best single frontier fixer solved 158, per new research posted to arXiv.

For teams paying per frontier-model call to fix code, that's a fifth of the cost per solve for the same hit rate. The extra scouting step adds under half a cent of GPU time per task.

Before any fix gets attempted, the small model reads the repository and writes a structured handoff describing the bug. Its reproduction claims get checked in a sandbox, with any false ones stripped out. That handoff, the task text, and the scout's hidden states feed a router that picks one of four frontier models to write the fix.

The researchers ran an ablation skipping the router entirely, always sending the handoff to the cheapest of the four fixers. It tied the routed system's score. The routing decision, the part that sounds like the innovation, isn't what's driving the result.

Calibration data in the paper backs this up. The handoff lifts solving ability for cheap models and slightly hurts the strongest one. That's what you'd expect if a good problem description matters more than which model reads it.

A separate paper in the same corpus argues coding-agent orchestrators need an explicit price list for each executor, not just access to one. Another straps a small failure monitor onto SWE agents to catch wasted effort before it piles up. The common thread: a cheap model doing verification or bookkeeping so the expensive model does less guessing.