CROQ improves LLM accuracy on multiple-choice and tool-selection tasks by using conformal prediction to build a coverage-guaranteed prediction set, pruning the answer choices to that set, and re-asking the narrowed question, since models answer better with fewer options. Its companion CP-OPT learns nonconformity scores that minimize prediction-set size while keeping coverage, outperforming raw logit scores. Learners study how distribution-free uncertainty sets, softmax-based scores, and set-size optimization combine to make LLM decision-making more reliable, evaluated on MMLU, ToolAlpaca, and TruthfulQA.
This ICML 2025 machine-learning paper improves the accuracy of large language models on multiple-choice decision tasks (MCQ answering, tool/API selection) by wrapping them in conformal prediction. Co…