ОЦЕНКА ЭФФЕКТИВНОСТИ СИНТЕЗА СТОХАСТИЧЕСКИХ МОДЕЛЕЙ С УПРАВЛЯЕМЫМИ СВОЙСТВАМИ

Abstract

This paper investigates the possibility of using large language models (LLMs) as sources of stochastic binary sequences with controllable statistical properties. The study demonstrates the temperature hyperparameter, which regulates the dispersion of the output token probability distribution, can be regarded as a tool for controlled entropy management of generated data. Based on the GPT-2 model, a method for converting probabilistic LLM outputs into fixed-length binary sequences is implemented and evaluated using entropy analysis and an adapted set of statistical tests analogous to NIST SP 800-22. The experiments reveal a monotonic increase in entropy with increasing temperature and a significant improvement in the statistical characteristics of the sequences. At temperatures T ≥ 2.0, up to 83.3% of the applied tests are passed, corresponding to the level of high-quality pseudorandom sequences for non-cryptographic applications. At the same time, a fundamental limitation of the approach is identified: weak but statistically significant periodic components persist due to the deterministic architecture of the model. The results define the applicability limits of LLMs in modeling, data synthesis, and algorithm testing tasks.