When are likely answers right? On Sequence Probability and Correctness in LLMs

Johannes Zenn, Jonas Geiping

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 leve…