What's this about?
Agents inspect working code, write behavioural specifications and tests, then filter tasks through repeated solution attempts. The resulting 5,545-task dataset trains coding agents without relying on issue histories.
What exactly is an executable verifier?
It runs a candidate solution against tests and returns a checkable reward. Hidden requirements or weak tests can still reward incorrect or overfitted solutions.
Why does this matter?
Author-reported preprint; 5 HF upvotes at fetch time.
- Existing repositories offer much broader task supply than curated issues.
- Agent-generated specifications need contamination and validity checks.
- Gains across five benchmarks are promising but require external replication.