What's this about?
RRSI evolves prompts, control flow, tools, memory and context management while constraining proposals and pruning brittle changes.
What exactly is regularisation?
It discourages a system from memorising its training tests. Here it limits edit bundles and rejects changes that are benchmark-specific, costly or no longer useful.
Why does this matter?
Author-reported preprint; 65 HF upvotes at fetch time.
- Self-improvement can overfit even when the backbone model is frozen.
- Out-of-distribution gains are smaller than in-distribution gains.
- A 30% token reduction suggests efficiency can be an explicit selection target.