- What changed
- Introduced CoVer, a framework using information-gain rewards and diversity-pruned tests to co-train language models as coders and verifiers, raising pass@1 by up to 7.1 points.
- Why you should care
- Information-gain rewards and diversity pruning mitigate permissiveness collapse and concentration bias in self-play code generation.
- Your move
- Watch. Monitor replication and adoption of CoVer techniques in broader code generation architectures.
- What to watch next
- Independent replication of CoVer benchmark results on alternative model backbones.
- Event
- research
- Event date
- Sep 21, 2026
- Relevant to
- General AI readers
