Position: Reasoning is a Learnable Rule-Based Process
Authors: Rachel Lawrence, Jacqueline Maasch
Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
Journal Reference: Proceedings of the 43rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026.
Submitted: 29 May 2026
Abstract
Autonomous reasoning is among the most scientifically and economically motivating topics in artificial intelligence (AI) today. Historically the purview of symbolic AI, recent advances have predominantly emerged from deep probabilistic generative models, such as large language models (LLMs). Despite immense interest and rapid progress, the generative AI community has not clearly converged on operational definitions for reasoning, and often implicitly rejects the historical treatment of this topic in logic and verifiable automated reasoning. This position paper contends that definitional ambiguity leaves the construct validity of reasoning evaluation unverifiable, undermining quantifiable progress toward trustworthy autonomous reasoning. We further argue that this ambiguity is addressable. To that end, we provide (1) operational definitions based on a synthesis of the literature, positioning valid and sound reasoning as a learnable rule-based process, and (2) a checklist for best practices in the communication of AI reasoning research.
1. Introduction
As of 2026, AI systems are increasingly deployed in domains that demand reliable reasoning, from scientific discovery to legal analysis and medical diagnostics. However, the notion of 'reasoning' in AI remains contested. While classical symbolic AI viewed reasoning as the manipulation of formal logic rules, the rise of deep probabilistic models has shifted the focus to learned statistical patterns. This shift has brought unprecedented capabilities, but also generated confusion about what constitutes genuine reasoning versus mere pattern matching. This position paper aims to clarify the discourse by proposing concrete, operational definitions of reasoning and offering a practical checklist for researchers to improve the clarity and validity of their work.
2. The Problem of Definitional Ambiguity
Despite its centrality, 'reasoning' lacks a universally accepted definition in the AI community. Modern generative models—especially LLMs—demonstrate impressive performance on tasks previously considered to require reasoning, such as mathematical problem solving and code generation. However, these systems are typically trained to predict the next token, not to follow explicit logical rules. This disconnect has led to a situation where evaluations of 'reasoning' often conflate different cognitive or computational processes, making it difficult to compare results across studies or to measure genuine progress. The result is a crisis of construct validity: we cannot reliably assess whether AI systems are truly reasoning or merely reproducing patterns from their training data.
3. Toward Operational Definitions
To address this ambiguity, we synthesize insights from logic, cognitive science, and AI to propose the following operational definitions:
- Reasoning: A rule-based process that produces conclusions from premises through valid inference steps. This process is learnable, meaning it can be acquired and refined through experience or training, and it is systematic, allowing for verification.
- Valid reasoning: A process where the conclusion necessarily follows from the premises, irrespective of the truth of the premises themselves. In formal terms, if the premises are true, the conclusion must be true.
- Sound reasoning: A valid reasoning process that also starts from true premises, thereby guaranteeing a true conclusion.
We position reasoning as a learnable rule-based process—a middle ground between the rigid rules of symbolic AI and the flexible pattern recognition of neural models. This definition allows AI systems to learn reasoning rules from data, while still providing a framework for verification and evaluation.
4. Implications for AI Research and Evaluation
Adopting these definitions has practical implications. First, it clarifies the goal: we aim to build systems that not only generate correct outputs but do so via verifiable inference steps. Second, it provides a basis for more rigorous evaluation: instead of merely measuring output accuracy, we can test whether a system’s internal process aligns with valid inference. Third, it encourages interdisciplinary collaboration, bridging the gap between neural and symbolic approaches.
5. Checklist for Best Practices in AI Reasoning Research
To facilitate clearer communication and more robust research, we propose the following checklist:
- Explicitly define the notion of reasoning used in your work, including the type of rules and inference steps involved.
- Ensure that your evaluation metrics track both outcome accuracy and process validity.
- Separate claims about validity vs. soundness, as they require different types of evidence.
- Document the training and inference mechanisms that give rise to reasoning behaviors.
- Compare against symbolic baselines where possible, to contextualize the role of learned rules.
- Report failure cases where the system appears to reason but does not, to highlight limitations.
- Use standardized datasets that include both valid and invalid reasoning examples to test generalizability.
- Reflect on the learning process: how are rules induced, and how are they updated with new information?
- Make your code and data publicly available to enable replication and scrutiny.
- Engage with philosophical and logical literature to ground your definitions in established theory.
6. Conclusion
Definitional clarity is not a mere academic exercise—it is essential for meaningful progress in AI reasoning research. By proposing operational definitions of reasoning as a learnable rule-based process, we provide a foundation for more reliable evaluation and more trustworthy systems. We call on the AI community to adopt these definitions and checklist to ensure that our advances are both measurable and interpretable.
Acknowledgments
The authors thank the anonymous reviewers for their constructive feedback, and the organizers of ICML 2026 for the opportunity to present this work.
References
[Full reference list omitted for brevity—includes works on symbolic reasoning, LLM reasoning, and evaluation methodologies.]
via ArXiv AI