When large language models try to show their work, things can go wrong. These models often improve at solving math or logic problems when they generate intermediate steps — a method known as Chain-of-Thought (CoT) prompting — but learning how to do that can make them rigid. They might do well on one kind of question and fall apart on others, or become overconfident in the wrong answers.
In “Soft Tokens, Hard Truths,” CDS Silver Professor of Computer Science, Mathematics, and Data Science Julia Kempe and collaborators propose a different way to train these reasoning steps. Rather than forcing the model to commit to one path of reasoning during training, they let it explore many possibilities at once using “soft” tokens — blurry combinations of multiple words or ideas instead of a single, fixed one. During inference, however, the CoT is generated in normal hard text.
