What's this about?
Each accepted code change becomes the agent edited in the next round. Hidden evaluations select improvements to search policy, memory and context management.
Is this unrestricted recursive self-improvement?
No. It is a bounded software optimisation loop with fixed tasks, evaluation infrastructure and acceptance rules—not an unconstrained model rewriting its own weights.
Why does this matter?
Author-reported preprint; 3 HF upvotes.
- Seven successive changes survived an eight-day autonomous run.
- Gains transferred to four held-out task families.
- Lower reward hacking is encouraging but was an unexpected correlation, not a guarantee.