protologue

Log Probabilities

Also called logprobs, token probabilities.

Log probabilities (logprobs) are the logarithms of the probabilities a model assigned to each token it generated, and optionally to the most likely alternatives, exposing how confident the model was at each step.

Description

Applications use logprobs to score classification labels, estimate confidence, flag likely hallucinations, rank candidate completions, and build autocomplete. They describe token-level likelihood, not factual correctness, so a fluent but wrong answer can still have high probability. Not every API or model exposes them.

Sources

  1. OpenAI (2023). Using logprobs.
  2. Kadavath et al. (2022). Language Models (Mostly) Know What They Know.

Cite this entry

Protologue. (2026). Log Probabilities. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0110). https://protologue.com/t/log-probabilities/

BibTeX
@misc{protologue_log_probabilities,
  title = {Log Probabilities},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0110},
  url = {https://protologue.com/t/log-probabilities/}
}

Markdown JSON