Research AlphaEvolve

Report

AlphaEvolve — Early Generations

Open AlphaEvolve infrastructure and scaffolds shipped; recorded evolutionary runs never progressed past early generations before the project paused.

Verdict Negative

Published
Jul 2025
Verdict
Negative
Verification
Verified
Question
What exists beyond the AlphaEvolve clone headline: runnable search infrastructure or demonstrated discovery?
Intent
Separate shipped evolutionary machinery from unearned claims about discovered programs.
Method
Audit public scaffolds, persistence machinery, evaluators, and recorded run depth before assigning the verdict.
Next test
Revive one benchmark with sealed lineage, evaluator fixtures, and a baseline before promoting any discovered program.

AlphaEvolve — Early Generations

Question

What exists beyond the AlphaEvolve clone headline: runnable search infrastructure or demonstrated discovery?

Context

AlphaEvolve is an open implementation of DeepMind’s evolutionary coding-agent pattern and the Club B propose/vary scaffold in the Recursive Self Improvement Lab: island-model populations, LLM mutation and crossover, sandboxed evaluation, and SQLite lineage. The quality bar is not whether the architecture is interesting; it is whether any recorded run discovered something worth promoting under an external evaluator the searcher cannot rewrite.

flowchart LR
  pop["Island population"]
  mutate["LLM mutate / crossover"]
  eval["Sandboxed evaluator"]
  persist["SQLite lineage"]

  pop --> mutate --> eval --> persist --> pop

Findings

  • The public engine and problem scaffolds exist: packing 26 circles in a unit square, 10-dimensional Rastrigin, kissing spheres, and small matrix multiplication.
  • Selection and persistence machinery (tournament / roulette / rank, SQLite generations) is in place so lineage can survive restarts.
  • Recorded runs never progressed past early generations before the project paused — no discovered-program benchmarks are claimed.
  • What publishes honestly is infrastructure and reference design; competitive results wait on a revival with sealed run cards.
  • For RSI Lab purposes this is a productive negative: evolutionary proposal machinery without held-out inheritance is not recursive self-improvement.

Method

LLM-driven evolutionary search under a sandboxed evaluator. Public surfaces are the open repository and strategy notes on elitism, selection pressure, and exploration. Verdict assignment required recorded run depth, not README aspiration.

Limits

  • No competitive discovered-program benchmark is claimed.
  • Early generations validate scaffolding, not search quality.
  • A revival needs a sealed run card, evaluator fixtures, a baseline, and a promotion gate independent of the mutating model.
  • Island-model activity is not archive-conditioned inheritance in the Wave-Causal F3/F4 sense.

Artifact

Repository: a3fckx/alphaevolve. No sealed notebook accompanies this status writeup.

  • evolutionary-computation
  • llm-agents
  • code-generation