Symbolic regression in practice: ~10k population needed even without agents

cephaloform · x · 2026-09-09

The author shares first-hand experience running symbolic regression with non-agent LLM generation (no reasoning, just emitting symbols): on hard problems you need a population size of 10k to reach truly strong solutions, and even with policy gradients on parameterized generative grammars at least 1k. They suggest a small (10M) AR model vomiting equations with at least 512 samples per step as an alternative.

Related event: Community Weighs In: Solving Hard Problems Needs 1K–10K Agents(2 posts)→

Original post →

More from Research

Research channel →