Position: Reasoning is a Learnable Rule-Based Process

Summary

This position paper argues that progress toward trustworthy autonomous reasoning in AI is hampered by persistent definitional ambiguity, particularly in work on large reasoning models derived from generative systems. While symbolic AI and formal logic have long treated reasoning as the application of explicit rules to derive valid and sound conclusions, recent literature on deep probabilistic models often omits or sidesteps such operational criteria. The authors contend that without shared, method-agnostic definitions, evaluations lack construct validity: benchmark accuracy on question-answering tasks cannot confirm that a model’s outputs result from rule-governed inference rather than memorization or superficial pattern matching.

To address this gap, the paper synthesizes concepts from logic, symbolic AI, and machine learning to define reasoning as a learnable process that applies exact rules to produce conclusions whose validity and soundness can be verified. The definition is presented in three complementary forms: natural-language statements for intuition, mathematical notation for precision, and pseudocode for implementation. Special cases such as logical deduction, Bayesian inference, reinforcement learning, and next-token prediction are examined to illustrate how the framework applies across paradigms. The authors introduce the notion of “reasoning zombies” to describe systems that emulate reasoning behavior without internal mechanisms that guarantee rule-based validity, drawing parallels to longstanding distinctions between emulation and genuine process in philosophy of mind and cognitive testing.

The work further identifies risks in current evaluation practices, including conflation of final-answer accuracy with the underlying reasoning process and insufficient attention to whether benchmarks actually measure the intended construct. It offers a concise communication checklist intended to improve clarity in research reporting, covering explicit operational definitions, distinctions between process and product, and acknowledgment of limitations in measuring autonomous reasoning. The authors position these contributions as prerequisites for measurable advancement toward artificial general intelligence, noting that reasoning is widely viewed as a necessary though not sufficient component.

Why it matters

Directly supports Dutch/EU priorities on ethical, transparent, and trustworthy AI by clarifying reasoning evaluation, which aids practitioners in building auditable systems compliant with regulations like the AI Act.