source: arXiv Artificial Intelligence: Position: Reasoning is a Learnable Rule-Based Process
level: research
A new position paper argues that the AI community has not agreed on what reasoning means. The authors say this ambiguity makes it impossible to verify whether systems truly reason. They propose operational definitions that treat valid and sound reasoning as a learnable rule-based process. The paper also includes a checklist for communicating AI reasoning research. The goal is to make progress toward trustworthy autonomous reasoning more measurable.
The paper notes that recent advances in reasoning have come mainly from deep probabilistic generative models. However, these models often implicitly reject the historical treatment of reasoning in logic and automated reasoning. The authors contend that without clear definitions, the construct validity of reasoning evaluations cannot be verified. This undermines quantifiable progress. They synthesize literature to define reasoning as a process that can be learned through rules, not just statistical patterns.
The checklist is intended to help researchers report their work in a way that allows others to assess whether reasoning is actually occurring. This matters for AI and data science because many systems claim reasoning abilities, but without shared definitions, those claims are hard to test. The paper does not present new experimental results. Instead, it offers a conceptual framework and communication standards. It is aimed at researchers who build or evaluate reasoning systems.
why it matters: Clear definitions and reporting standards help AI practitioners verify whether a system truly reasons, reducing the risk of overclaiming capabilities.
source: arXiv Artificial Intelligence: Position: Reasoning is a Learnable Rule-Based Process