ICML paper argues reasoning is a learnable rule-based process
Researchers propose operational definitions to fix AI reasoning's verification crisis
A new position paper from Rachel Lawrence and Jacqueline Maasch, accepted at ICML 2026 in Seoul, confronts a core problem in AI reasoning research: nobody agrees on what reasoning actually is. Published on arXiv (2608.12325), the paper contends that the generative AI community has largely abandoned the formal definitions developed in symbolic AI, logic, and verifiable automated reasoning. This definitional ambiguity means current benchmarks for reasoning may not measure what they claim, leaving the construct validity of reasoning evaluation effectively unverifiable.
To address this, the authors provide two main contributions. First, they synthesize the literature to offer operational definitions that position valid and sound reasoning as a learnable rule-based process — moving the concept away from nebulous probabilistic behavior and toward something more structured and testable. Second, they introduce a checklist of best practices for communicating AI reasoning research, aiming to standardize how models are evaluated and compared. The paper is a call for the field to converge on shared terminology, which the authors argue is essential for quantifiable, trustworthy progress in autonomous reasoning.
- Paper argues current generative AI reasoning evaluations lack construct validity due to ambiguous definitions
- Authors synthesize literature to define reasoning as a learnable rule-based process with valid and sound reasoning
- Provides a best-practices checklist for communicating AI reasoning research, accepted at ICML 2026
Why It Matters
Standardizing reasoning definitions enables verifiable benchmarks, accelerating trust in autonomous AI systems.