From Paper to Program: Knowledge Externalization and Bottleneck Diagnosis in AI-Assisted Quantum Many-Body Programming
arXiv:2604.04089
The paper studies how large language models can be guided to translate quantum many‑body research papers into working code by explicitly externalizing tacit knowledge such as gauge choices and sign conventions, and it evaluates a human‑in‑the‑loop workflow on DMRG and Pfaffian‑MPS tasks.
Abstract
Implementing quantum many-body algorithms from the literature is fragile when executable correctness depends on tacit choices of index order, gauge, fermionic sign, contraction sequence, and scaling strategy. We formulate this as a knowledge-externalization problem and test a staged, human-in-the-loop workflow that converts source papers into reviewed technical specifications before code generation. DMRG from Schollw{ö}ck's review serves as calibration: specification-guided implementations pass all 16 model pairings, compared with 6/13 direct attempts, while a prose-specification ablation retains the improvement when externalized content is preserved without \LaTeX{} form. Pfaffian conversion of HFB states to MPS from a five-page Letter provides a closed-world stress test using standalone NumPy/SciPy/Matplotlib implementations without supplied tensor-network code. The workflow yields 11/26 audited passes, versus none under direct prompting. Cross-specification transfer is asymmetric: GPT~5.5 implements four non-GPT specifications successfully, whereas the reverse direction fails in four tested cases. These results support two distinguishable bottlenecks: paper-to-code ambiguity, reduced by explicit specification, and residual implementation-agent capability. Iterative meta-specification shifts but does not remove the latter. The resulting \emph{Paper-to-Program Many-Body} protocol couples expert review, provenance checks, production-scale gates, and physics oracles to provide an auditable pathway from published many-body theory to validated code.
20 pages, 4 figures. Substantially revised presentation, related-work context, reproducibility documentation, validation-gate wording, limitations, and generative-AI disclosure; numerical results unchanged