Token probability is the normalized probability a language model assigns to each possible next token at a particular position. The model distributes probability mass across its vocabulary after converting logits with softmax; reading the probability of each observed next token measures how compatible a proposed sequence is with the model's preferences. section 5.4
Compute token probabilities for a proposed sequence as follows:
- Multiply the conditional probabilities to estimate the sequence's joint probability; very small products can make direct comparison numerically inconvenient. section 5.5
Berlin versus Bridge
For "The capital of Germany is Berlin," the displayed next-token probabilities produce a joint probability of 5.9372e-08. Replacing Berlin with Bridge lowers the final token probability to 2.9802e-07 and the joint probability to 1.0481e-13, indicating much weaker model support.
Token probabilities support response scoring: consistently high values suggest model compatibility, while many low values suggest weak support. They can rank candidate responses, break ties in self-consistency, and provide a signal for iterative self-refinement. section 5.4 section 5.1
