Researchers propose operational definitions for AI reasoning as a learnable rule-based process

A new position paper on arXiv argues that the AI community lacks operational definitions for reasoning, creating ambiguity that undermines the construct validity of evaluations and slows progress toward trustworthy autonomous reasoning systems. The authors, Rachel Lawrence and Jacqueline Maasch, contend that this ambiguity stems from a disconnect between modern generative AI approaches and the historical, logic-based treatment of reasoning in symbolic AI. They propose that reasoning be treated as a learnable rule-based process, offering operational definitions grounded in a synthesis of prior literature.

To address the communication and evaluation challenges arising from definitional ambiguity, the paper introduces a checklist of best practices for AI reasoning research. The authors assert that clearer definitions and standardized reporting would improve the verifiability of reasoning evaluations, enabling more quantifiable progress in the field. The proposal is situated within broader discussions about the reproducibility and trustworthiness of AI systems, particularly as autonomous reasoning becomes a central focus of both scientific and economic interest.

The paper was submitted to the 43rd International Conference on Machine Learning (ICML 2026) and is available as arXiv:2608.12325. While the authors do not claim empirical results, they frame their contribution as a position advocating for definitional clarity and methodological rigor in reasoning research.