IBM, NIST/UMD, MIT preprint reports 699 magic-state factory classes found with LLM agents
Researchers from IBM Quantum, QuICS (NIST/UMD) and MIT describe a neuro-symbolic search framework that pairs large language model agents with deterministic solvers to explore magic-state distillation protocols. In a preprint titled "Symmetry and AI-assisted discovery of magic-state factories" (not yet peer reviewed), they report 699 factory classes, 564 of them previously unknown. Reported highlights include a [[63, 11, 3]] code yielding 11 pure T states at 5.73 inputs per output (overhead exponent γ = 1.589), a [[850, 128, 6]] code with γ = 1.057, and a [[1715, 287, 6]] code with γ = 0.998, which the authors describe as the smallest known pure-T factory with γ below 1. The team also found factories outputting entangled states such as CCZ⊗2 and T ⊗ CCZ⊗4, and released the verified circuits in an open-access repository.