Concretized Proposition Prompting Resolves Composition-Knowledge Dichotomy in Large Language Models
06:00 · July 11, 2026 · arXiv cs.AI RSS
LLMs often struggle to balance compositionality with knowledgeability, a challenge we define as Composition-Knowledge Dichotomy. To address this, we propose Concretized Proposition Prompting (CPP), a framework that explicitly concretizes propositions relevant to questions. The results demonstrate that CPP significantly enhances reasoning performance, particularly in medical benchmarks where precise knowledge is paramount, while being competitive on math benchmarks where deductive reasoning is prioritized. Additional experiments reveal that CPP is scalable to various foundation models and parameter sizes, being a fundamental paradigm that bridges the gap between composition- and knowledge-based approaches. Consequently, CPP resolves the composition-knowledge dichotomy by providing a solid foundation for logically organized and factually grounded reasoning.
Summary
Concretized Proposition Prompting (CPP) addresses a core limitation in large language models: the Composition-Knowledge Dichotomy. This refers to the tension between compositionality, which emphasizes structured logical reasoning through intermediate steps, and knowledgeability, which stresses the retrieval and application of accurate factual content. Existing techniques such as chain-of-thought prompting and its variants tend to favor one axis over the other, leaving models prone either to post-hoc rationalizations that lack truth-value checks or to deductions that fail to ground statements in verifiable propositions.
CPP counters this polarization by inserting an explicit proposition-generation stage before the final answer. The model first produces question-relevant statements and classifies them into four categories—true-positive, true-negative, false-positive, and false-negative—according to whether each affirms or negates a fact or a fallacy. These categorized propositions are then supplied to a downstream answer model, which uses them as concrete anchors for its reasoning. The approach therefore combines the logical organization of composition-based methods with the factual grounding required by knowledge-based methods.
Experiments across eight question-answering datasets spanning commonsense, mathematical, and medical domains show that CPP improves accuracy relative to prompting baselines that focus on either structure or evidence alone. Gains are most pronounced on medical benchmarks, where precise factual discrimination is essential, while performance on mathematical tasks remains competitive with methods optimized for deductive chains. The framework demonstrates consistent benefits when applied to multiple open-source foundation models, including Llama, Qwen, Phi, Gemma, and Mistral families, and maintains its advantages across parameter scales from 7B to 72B.
Why it matters
CPP offers Dutch AI researchers and practitioners a robust, scalable prompting technique to enhance LLM reliability and reduce hallucinations. This aligns perfectly with the EU's strict requirements for trustworthy, transparent, and factually grounded AI systems, especially in high-stakes domains like healthcare.
