V.N. Dobrynin1, A.A. Milovidova2, I.A. Filozova3
1–3 State University «Dubna», Institute of Systems Analysis and Management (Dubna, Russia)
2 MIREA – Russian Technological University, Institute of Information Technologies (Moscow, Russia)
3 Joint Institute for Nuclear Research, Meshcheryakov Laboratory of Information Technologies (Dubna, Russia)

1 arbatsolo@yandex.ru

Designing reproducible operational strategies when using LLMs presents challenges, and unstructured prompting leads to inconsistent results. This paper presents a methodology for normalizing semantic code – transforming it from a working notation into a rigorous, executable grammatical form.

Formal foundations for a semantic coding language and an LLM architecture serving as its interpreter have been developed. However, a gap exists between the conceptually clear working notation and the rigorous grammatical form. Using an example from research administration (processing a publication list from an institutional repository), the study demonstrates the methodology for normalizing semantic code from an initial entry to a grammatically rigorous form. Common deficiencies in working entries, such as mixed levels of abstraction, incomplete operators, unstable ontological frameworks, implicit dependencies, and incomplete synthesis, have been identified, and methods for resolving them have been proposed.

The outcome is normalized semantic code suitable for interpretation. This methodology can be applied to the design of reproducible semantic programs not only in the fields of scientometrics and research management but also across a much broader range of tasks.

Dobrynin V.N., Milovidova A.A., Filozova I.A. Reproducible semantic computations: an experiment in semantic program normalization // Neurocomputers. 2026. V. 28. № 4. P. 68–74. DOI: https://doi.org/10.18127/j19998554-202604-06

1. Zamfirescu-Pereira J.D., Wong R.Y., Hartmann B., Yang Q. Why Johnny can't prompt: How non-AI experts try (and fail) to design LLM prompts. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 2023. P. 1–11.
2. Sahoo P., Singh A.K., Saha S. et al. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint. 2024.
3. Liu Y., Li D., Wang K. et al. Are LLMs good at structured outputs? A benchmark for evaluating structured output capabilities in LLMs. Information Processing & Management. 2024. V. 61. № 5. Id. 103809.
4. Yao S., Zhao J., Yu D. et al. ReAct: Synergizing reasoning and acting in language models. ICLR. 2023.
5. Geng S., Josifoski M., Peyrard M. et al. Grammar-constrained decoding for structured NLP tasks without finetuning. EMNLP. 2023. P. 1–15.
6. Filozova I.A. Institutsional'nyj repozitorij publikatsij kak komponent tsifrovoj ekosistemy krupnoj nauchnoj organizatsii. Diss. … kand. tekhn. nauk. SPb. 2025. (in Russian)