Skip to content
BENIBIX Learn

LLM Non-Determinism

AI💬 LLMsText Generation

Also known as Non-Deterministic Output (LLM), LLM Nondeterminism

The same prompt can give a new answer each time you ask.

Think of it like

It is like asking a friend the same riddle twice. Most days you hear the same answer, but not always.

Example

Emma sends one prompt twice at temperature 0 and gets two answers. They match until the third sentence, then go different ways.

Why it matters

Tests that expect the exact same text will break. Checks need to judge the meaning, not the exact words.

Definition

LLM non-determinism means the same prompt can give different answers, even with the same settings. Most of it comes from sampling, which picks words with some chance on purpose. But even at temperature 0, hosted models often do not give the same answer twice. How busy the server is changes how requests are grouped, and that can shift the math just enough to tip a close choice. Once one word differs, the rest of the answer can drift.
Two runs at temperature 0 match until one close choice tips the other way.

Common mistake

Thinking temperature 0 always gives the same answer. Providers say it does not.

Key takeaways

  1. 1
    Sampling adds chance. On purpose.
  2. 2
    Temperature 0 is no lock. Server math still varies.
  3. 3
    One word drifts. The rest can follow.

Learn it properly