Proceedings of the Annual Meeting of the Cognitive Science Society

Human reasoning is characterized by the rational use of resources to optimize performance under constraints. Recently, inference-time scaling has improved the reasoning performance of Large Language Models by increasing test-time computation. Instruction-tuned (IT) models explicitly generate long reasoning traces, whereas Large Reasoning Models (LRMs) are trained via reinforcement learning to discover reasoning paths that maximize accuracy. However, it remains unclear whether resource-rationality can emerge from such scaling without explicit rewards related to computational costs. We introduce a Variable Attribution Task (VAT) in which models infer which variables determine outcomes given candidate variables, input–output trials, and predefined logical functions. By varying the number of candidate variables and trials, we systematically manipulate task complexity. Both models exhibit a transition from brute-force to analytic strategies as complexity increases. IT models degrade on XOR and XNOR functions, whereas LRMs remain robust. These results suggest that resource rationality can emerge from inference-time scaling itself, even without explicit cost-based rewards.