The honest part: the ablation is mixed. Searching for distance alone cut similarity slightly more, and random directions passed the novelty check more often, so the filtering and prompting stages do a lot of the work. There is no human evaluation and no temperature baseline yet. Turning a well-chosen gap into a better answer, not just a different one, is the open problem.

Caveats: agents must share one model; AIME has 30 problems, so one problem is 3.3 points; two of the four papers are preprints. For speeding up shared context at the systems level, see KVCOMM by @HankYe1 and colleagues.

Related: RRSI by @richardxp888 et al. finds regularizing the RSI loop improves out-of-distribution generalization, and Misevolve by @Amanda_JShao et al. shows safety can drift as memory accumulates. Caveat for ours: a few evolution stages, not open-ended RSI, and gains in the skill and model channels are small.