arXiv:2610.06623
Large Language Model-Guided Discovery of Weight-Five Bivariate Bicycle Codes
Juan Cruz-Benito
Abstract
Building on our earlier program-evolution workflow guided by large language models (LLMs), we study weight-five bivariate bicycle (BB) and perturbed bivariate bicycle (PBB) codes. The resulting catalogue contains 1,142 distinct code proposals, including 1,081 nonbaseline proposals attributable to LLM-generated programs. Across the catalogue, we certify connected Calderbank--Shor--Steane (CSS) realizations [[96,4,10]], [[140,6,10]], and [[180,4,14]]. A post-search comparison certifies seven imported Lin--Pryadko archive constructions. For leading parameter triples also represented in that archive, we provide exact distance evidence, explicit bivariate presentations, and verified component reductions. A basis-independent connectivity analysis identifies 409 of the 1,142 catalogue entries as disconnected and shows that 73.1\% of the classes with exact distance certificates contain repeated connected components. Algebraic analysis organizes the connected CSS classes into order-3, order-7, and order-15 cyclotomic-kernel strata. The strongest exact connected PBB parameter point is [[216,4,10]], attained by two distinct component classes. Among the 936 distinct CSS proposals from the LLM-guided campaign with a recorded positive distance, 816 (87.18\%) are certified at $d\geq5$. For comparison, three random-search controls each sample 6,444 CSS code proposals uniformly without replacement, using the same per-lattice and encoded-dimension sample counts as the LLM-guided campaign. In these controls, 4,672--4,785 proposals (72.50--74.26\%) meet the same criterion. The LLM-guided campaign has the higher certified yield, while the random controls cover more connected classes. Together, these results extend LLM-guided discovery to a more constrained code family and provide a reproducible structural and exact-distance account of its strongest candidates.
AI Summary
The paper studies whether LLM-guided program evolution can find weight-five bivariate bicycle (BB) and support-contained perturbed BB (PBB) codes, and characterizes their distances, connectivity, and algebraic structure. Four searches produce a retained catalogue of 1,142 proposals, including 1,081 attributed to LLM-generated programs, grouped into 630 presentation classes. Distance evidence combines decoder estimates, exhaustive low-weight searches, verified logical witnesses, and MILP, with exact status requiring matching rigorous bounds. Connected CSS realizations include $[[96,4,10]]$, $[[140,6,10]]$, and $[[180,4,14]]$; the last has the largest exact connected CSS catalogue value of $kd^2/n=4.356$. These parameter triples already occur in the Lin–Pryadko archive; equivalence is established for $[[96,4,10]]$. Seven archive constructions receive exact certificates, including $[[132,4,12]]$ with $kd^2/n=4.364$. The strongest exact connected PBB figure of merit is 1.85, attained by two presentation classes with parameters $[[216,4,10]]$. Connectivity analysis finds 409 proposals disconnected. Sampled connected CSS classes are organized into order-3, order-7, and order-15 cyclotomic-kernel strata. Among 936 distinct logged CSS proposals, 87.18% have certified $d\geq5$, versus 72.50–74.26% in three stratified-uniform controls. Controls cover more connected classes and each samples the $[[180,4,14]]$ class. The control comparison does not establish attempted-call or compute efficiency: campaign failures are incompletely logged, repetitions differ, and high-distance follow-up is asymmetric. Presentation classes are not full code-equivalence classes, most PBB local-Clifford classifications remain unresolved, and circuit-level performance is not evaluated.
An ensemble of Claude Opus 5 and GPT-5.6 Sol generated program variants, with model labels retained as provenance.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.