Log Probability (logprob)

The logarithm of the probability a model assigns to a specific token given the preceding tokens. Because probabilities lie between zero and one, a logprob is less than or equal to zero: values closer to zero indicate more likely tokens, and exp(logprob) recovers the corresponding probability. Summing logprobs scores a sequence without multiplying many small numbers.

Entropy summarizes uncertainty across the full distribution; logprob observes the mass assigned to one realized or candidate token; perplexity derives from average negative log-likelihood across a sequence. Logprobs help inspect alternatives, compare answers under the same protocol, and support abstention, but they do not represent calibrated confidence. Tokenization, length, model, and inference configuration must remain controlled when comparing values.