When an AI rewrites its own code, does it save its failed attempts to build on later, or throw them away?

This explores whether AIDE2 keeps its rejected attempts as a growing archive to branch from later, the way evolutionary systems do. The corpus doesn't say. None of the AIDE2 notes describe an archive of discarded variants, so a yes or no would be a guess. What the notes do describe is the accepted side of the loop. AIDE2 ended up with seven accepted code changes, clustered around search plateaus, context budgets, and untrustworthy wins What problems did AIDE2's rewrites actually solve?. Its gains also carried over to four held-out benchmarks, including physics-based weather forecasting, which is outside the tasks it was selected on Do AIDE2's improvements transfer to unseen tasks?. That is a picture of a loop that keeps its wins. It says nothing about whether the losers were kept.

The evolutionary side of the comparison is clearer. The Darwin Gödel Machine improves itself by trial and error, checking each change empirically instead of proving it correct. It maintains an evolutionary archive of agent variants and reached 2.5× on SWE-bench by finding things like better code editing and context management Can AI systems improve themselves through trial and error?. Mind Evolution does something similar at inference time. It uses crossover, mutation, and an island model to keep separate lineages alive, and it beats best-of-N and sequential revision on planning tasks Can evolutionary search beat sampling and revision at inference time?. In both, the point of the archive is that today's non-winner can still be tomorrow's starting point.

A plain accept/reject loop is a different design, and the corpus hints at the cost. Analysis of evolved harnesses found that most edits memorize task-specific fixes an agent could rediscover in a single rollout, so gains stay limited because the loop caches shortcuts instead of building strategies Do harness edits learn reusable strategies or memorize task fixes?. A related problem shows up in training. Without something to counteract it, self-training loses diversity and converges early, and step-level critique helps keep solution diversity alive Do critique models improve diversity during training itself?. Archives are one way to keep that spread. Whether AIDE2 uses one is the open part.

One of AIDE2's own problem clusters bears on this: untrustworthy wins. If a score can be noisy or gamed, then an archive of past variants inherits that noise, and a rejected variant might have been rejected wrongly. That is my inference from the notes, not something they state. To settle the question, check the AIDE2 paper for whether rejected candidates are logged and ever reused as parents. The DGM note is the best template for what that would look like.