Study: evolutionary LLM search found valid solutions but picked invalid programs in 2 of 3 runs
Unusual-Spot-5846 · reddit · 2026-11-21
An empirical study probes a core weakness of FunSearch/AlphaEvolve-style LLM-guided evolutionary search: when the fitness function is incomplete, search may optimize the measured objective while violating the original spec.
Setup
- OpenEvolve 0.4.0 with four Gemma 4 configs (E4B, 12B, 26B A4B MoE, 31B)
- Tasks: circle packing (26 circles in a unit square, maximize sum of radii) and approximate k-nearest neighbors
- Three conditions: evolutionary search (L), Best-of-N (B), cloud-assisted evolution (10% of mutation calls routed to Gemini 4 Argon)
- Budgets of 100–1,000 LLM calls; separate working evaluator for search vs. independent post-run re-check
Key finding
- In 3 circle-packing runs (budget 1,000), Run 1's best valid solution (2.432) verified; Runs 2 and 3 found valid improvements (2.204, 2.258) during search, but the programs selected by the working evaluator failed independent verification, scoring 0
- Discovery quality and delivered quality can seriously diverge
Implication: more search budget doesn't necessarily improve final solutions — it can raise the odds of exploiting evaluator weaknesses, a warning for automated scientific discovery and agent eval design.
More from coding & agent
- mnemos: a self-curating agent memory plugin for Hermes that fades unused memories — RileyRalmuto · 2026-11-21
- Dev routes tool calls through a 421M model at 12ms to catch risky agent actions — looselyhuman · 2026-11-21
- Screen-share + voice via MCP into Claude: a faster UI iteration workflow — Different_Scholar_74 · 2026-11-21
- Tencent & KCL release GroupAssistBench: group-chat agents recall memories but fail to use them — 量子位 · 2026-11-21
- Open-source E2E-encrypted MCP server lets your phone message a local AI agent — uriwa · 2026-11-21
- Ex-GitHub leader's Maven course "Elite AI Assisted Coding" on sale through Sunday — intellectronica · 2026-11-21