Hidden Nondeterminism in AI

Large Language Models (LLMs) are often expected to produce reproducible results when the same prompt and settings are used. However, new research shows that there can be subtle differences inside these models, even when their final answers appear to be exactly the same.

Researchers from the Information Processing and Telecommunications Center (IPTC) of the Universidad Politécnica de Madrid (UPM) have investigated this hidden nondeterminism by analyzing the probability assigned to individual tokens (the small units of text that LLMs use to generate their responses). Their work, published in IEEE Transactions on Computers, shows that these probabilities can fluctuate between runs due to microscopic numerical differences arising from the parallel execution of computations on GPUs.

The study found that this effect is particularly relevant when the model is uncertain about which token to select. Highly confident choices tend to remain stable, while intermediate-probability decisions are more sensitive to small variations. The researchers also observed that larger inference batch sizes can increase this probability of variation. These patterns were found across different models and GPU platforms, suggesting that this is not an isolated behavior of a particular system.

One of the most important findings is that an LLM may generate exactly the same text on different runs while its internal probability distribution is changing. This “illusion of consistency” could hide potential fragility that cannot be detected simply by comparing the final answers.

Understanding this hidden behavior could help improve AI benchmarking, make comparisons between models more reliable, and support the development of more robust and trustworthy AI systems. It could also be valuable for applications where reproducibility and reliability are particularly important, such as AI evaluation, production systems and safety-critical environments. The researchers’ additional material highlights the relevance of these findings for accurate benchmarks, fair model comparisons and reliable production pipelines.

Bibliographic reference:

Fu, T., Martínez, G., Conde, J., Arriaga, C., Qi, X., Liu, Shashan, L. & Reviriego, P. Beyond Reproducibility: Token Probabilities Expose Large Language Model Nondeterminism in IEEE Transactions on Computers, Early Access.  https://doi.org/10.1109/TC.2026.3732142

Authors:

Gonzalo Martínez Ruiz de Arcaute: ORCID/ LinkedIn

Javier Conde Díaz: GS /ORCID/ LinkedIn

Carlos Arriaga Prieto: GS / ORCID/ LinkedIn

Pedro Reviriego Vasallo: GS / ORCID/ LinkedIn

LinkedIn: https://www.linkedin.com/company/iptc-upm/

For more information: www.iptc.upm.es

Share this: