Concretized Proposition Prompting Resolves the Composition-Knowledge Dichotomy in Large Language Models

This paper introduces Concretized Proposition Prompting (CPP), a framework that explicitly concretizes propositions relevant to a question before reasoning. The authors frame the tension between compositional reasoning and factual knowledge as the “Composition-Knowledge Dichotomy.” According to the abstract, CPP substantially improves reasoning performance on medical benchmarks, where precise knowledge is important, while remaining competitive on mathematics benchmarks emphasizing deduction. Additional experiments reportedly show scalability across foundation models and parameter sizes. The supplied material is limited to the abstract, so benchmark names, baselines, effect sizes, implementation details, and statistical evidence cannot yet be independently assessed.