Nondeterminism in LLM Outputs

A language model generates tokens from a probability distribution. Sampling settings such as temperature and top-p can make output variation expected, but setting temperature to zero does not guarantee identical responses in every hosted or distributed serving system. Small numerical differences, parallel execution, model updates, routing, and tool results can change a token choice and lead to different later text. Some API providers expose a seed parameter and backend fingerprint to improve reproducibility. OpenAI’s documentation describes the seed as best effort and recommends checking the system fingerprint; even when request parameters and fingerprint match, outputs may still differ. Pinning model snapshots, keeping prompts and request settings fixed, and recording tool versions can make comparisons more interpretable, but does not create a universal determinism guarantee. Repeated-run variation matters for tests, caching, debugging, and user-facing behavior. For an evaluation, either control randomness where supported or run multiple samples and report variability. Use semantic or structured assertions when exact text matching is too brittle. Cache only when application semantics allow it, and do not rely on a seed as a security or correctness mechanism. Reproducibility requires recording more than a prompt: model identifier, seed, temperature, top-p, system fingerprint, tool outputs, code, and relevant runtime configuration. Some providers do not expose all of these fields. Treat exact repeatability as a property to measure under a documented setup rather than an assumption based on a single parameter.