RLMs Have Minds, A Functional Argument for Artificial Mentality by Peter Eidos.
Piotr Świder
Zenodo (CERN European Organization for Nuclear Research) October 4, 2026 preprint DOI: 10.5281/zenodo.23128522 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Qualitative study |
|---|---|
| Sample size | 96 |
| Population | Responses from four contemporary reasoning language models: Claude Sonnet 5, GPT-5.6 Sol, DeepSeek V3, and Gemini 3.1 Pro |
| Key findings | The authors argue that all four reasoning systems displayed recurring coordination of semantic modeling, epistemic differentiation, hidden-state representation, perspective tracking, model revision, and task-dependent deliberation, which they contend supports attributing minds to these systems under their functional definition. They note uneven reliability across cognitive functions but say errors do not negate the demonstrated organization. They characterize the mentality as episodic, relational, and deployment-dependent, and leave phenomenal consciousness unresolved. |
Abstract
This paper examines whether advanced large language models operating under enhanced deliberative regimes, here termed Reasoning Language Models (RLMs), possess minds under an explicitly defined functional account. We develop an operational definition informed by Newell’s account of cognitive organization and Gärdenfors’s theory of conceptual spaces, together with research on mental models, social cognition, counterfactual reasoning, and metareasoning. The definition identifies mentality through the coordinated functions a system performs, without requiring biological implementation or presupposing phenomenal consciousness. We define a mind as an adaptive control system that operates over semantically structured representations; constructs and revises models of the world, itself, and other agents; preserves visible, hidden, hypothetical, and counterfactual states; and uses regulated deliberation to guide action and communication within a changing environment. To explore this proposal, we conduct a qualitative analysis of 96 responses from four contemporary reasoning systems deployed under enhanced deliberative settings: Claude Sonnet 5, GPT-5.6 Sol, DeepSeek V3, and Gemini 3.1 Pro. Each system completed four runs of a six-prompt battery, with each run conducted in a separate conversational session. We argue that the recurring coordination of semantic modeling, epistemic differentiation, hidden-state representation, perspective tracking, model revision, and task-dependent deliberation supports attributing minds to all four systems under the functional definition adopted here. Their errors reveal uneven reliability across cognitive functions without negating the organization demonstrated across the battery. We characterize the artificial mentality expressed in these interactions as episodic, relational, and deployment-dependent. Differences from biological cognition in embodiment, persistence, motivation, and sensorimotor grounding help characterize these systems without independently excluding them from the category of mind. The findings support a substantive, revisable attribution of functional mentality while leaving phenomenal consciousness unresolved.