Quantum Error Correction Meets AI for Science: Lessons from LLM-Guided Code Discovery

Abstract

Invited keynote at the AI for Circuit Synthesis, Optimization, and Discovery (AI4QC) workshop, held alongside IEEE Quantum Week 2026 in Toronto. The talk places LLM-guided program evolution within the broader ‘AI for science’ landscape (AlphaFold, GNoME, FunSearch, AlphaEvolve) and walks through a verifier-guided evolutionary pipeline, built on OpenEvolve, that mutates Python programs generating bivariate-bicycle and perturbed bivariate-bicycle quantum LDPC codes. Screening roughly 2×10^5 candidates across five campaigns surfaced 465 BLISS-distinct codes (97 CSS, 368 non-CSS) at block length n ≤ 360, backed by an independent verification stack — GF(2) rank computation, MILP-based distance bounds, BLISS Tanner-graph deduplication, and local-Clifford equivalence checks — built to expose and correct the search’s own blind spots. The talk closes with transferable lessons on where LLM-guided evolution complements, rather than replaces, reinforcement-learning approaches to quantum error-correcting code discovery.