A probabilistic language model estimates how likely tokens or sequences are given preceding context. Modern autoregressive models generate by repeatedly calculating a distribution over the next token, selecting a continuation, and feeding it back into the context for the following step.
These token probabilities support generation but do not directly measure whether a complete statement is true. Several confident-looking sentences can be assembled from locally probable tokens, and answer-level uncertainty may require separate estimation, retrieval, verification, or calibration methods.
Acronyms and aliases
probabilistic token prediction variant
Related terms
Frequently asked questions
What does a probabilistic language model predict?
It predicts a probability distribution over possible tokens or sequences conditioned on the context it has already received.
Are token probabilities the same as factual confidence?
No. They describe likely continuations under the model, not a calibrated probability that the generated claim is factually correct.
