LLM-guided program evolution improves 10 best-known circle-packing solutions (Packomania csqv, N=101-114) [R]

I used an LLM to iteratively evolve an optimization algorithm rather than solve the packing directly. Starting from a simple seed solver, the LLM proposes algorithmic changes guided by a scoreboard of results and a history of prior attempts, and each candidate is scored by an independent verifier so improvements are kept and failures discarded. On the Packomania csqv benchmark it improved the best-known sum-of-radii for 10 values of N from 101 to 114, by 2.4 to 5.4%, in 15 iterations. Total LLM cost was $27.72. Packomania accepted the results independently.

Paper: arxiv.org/abs/2609.05093

Code + solutions: github.com/ucsandman/discovery-loop

Benchmark: packomania.com/csqv/csqv.html

Happy to discuss the plateau-detection stopping rule, that's the piece I'd most want critique on.