LLM-based frameworks for algorithm evolution
Autonomous Claude Code agent that iteratively modifies code, runs evaluations, and logs experiments in a continuous research loop. Uses Claude's built-in tool use to edit files, run commands, and reason about results.
AdaEvolve search from the SkyDiscover framework. Adaptive multi-island evolutionary search with UCB-based island selection, periodic migration, unified archive of historical winners, and paradigm-breakthrough prompts triggered when global progress stalls.
Beam search from the SkyDiscover framework. Maintains a fixed-width beam (default 5) and balances quality and diversity when picking parents for the next round (diversity_weighted strategy).
Best-of-N from the SkyDiscover framework. Generates N independent candidates from the seed (no evolution between candidates) and reports the highest-scoring one. A pure exploration baseline.
Top-K beam search from the SkyDiscover framework. Maintains a sorted population of the K best programs by score and samples inspirations from it for each new candidate. The simplest baseline among SkyDiscover's search algorithms — no islands, no adaptation.