When are likely answers right? On Sequence Probability and Correctness in LLMs
This work clarifies the limitations of using sequence probability as a proxy for correctness in LLM decoding, providing practical guidance for practitioners in decoding, self-consistency, and self-improvement.
The paper investigates the relationship between sequence probability and correctness in LLMs across decoding methods, hyperparameters, and prompt-answer pairs. It finds that while higher sequence probability correlates with correctness across pairs within a dataset, it does not reliably predict correctness when changing hyperparameters or methods, nor for repeated responses to the same prompt.
Many decoding methods for large language models can be understood as shifting probability mass toward outputs that are more likely under the model, either locally at the token level or globally at the sequence level. Therefore, their success depends on a fundamental question: when does sequence probability, that is, the conditional probability of a continuation given a prompt, actually align with correctness? In this paper, we set out to quantify this relationship across decoding methods, models, and benchmarks at four levels: across decoding methods, across hyperparameters within a method, across prompt-answer pairs within a dataset, and across repeated responses to the same prompt. We find that higher sequence probability is often predictive of correctness across prompt-answer pairs within a fixed dataset. However, this relationship does not generally transfer to decoding decisions: increasing sequence probability by changing hyperparameters or methods does not reliably improve accuracy. Further, sequence probability is not a good indicator of correctness for responses to the same prompt. These findings clarify when decoding can and cannot be expected to improve correctness, and provide practical guidance for decoding, self-consistency, and verifier-free self-improvement.