When Humans Stop Thinking, Why Do Reasoning Models Persist? — New Findings on the Automated Allocation of Cognitive Deliberation

Latest research reveals the core differences between humans and reasoning models in the allocation of deliberation (careful thinking): when humans strategically give up deep thinking due to increased task difficulty, reasoning systems based on large language models still maintain continuous cognitive engagement. This study was conducted by the henryhyw team, titled "Humans Disengage, Reasoning Models Persist", and has been submitted to Computational Brain & Behavior, marking the first systematic revelation of the essential differences in cognitive resource allocation mechanisms between the two.

This study focuses on the core question: How do humans and reasoning systems based on large language models allocate cognitive resources when facing tasks that require deep thinking? Through a carefully designed experimental paradigm, the study aims to reveal the essential differences in deliberation allocation mechanisms between the two.

The research team adopted the Item-Controlled Dissociation paradigm. By controlling the cognitive load and complexity of tasks, they observed behavioral differences between human participants and reasoning models. The key lies in independently manipulating the objective difficulty and subjective perceived difficulty of tasks to isolate the true deliberation allocation strategies.

When task difficulty increases to a certain level, human participants tend to "strategically give up"—reducing cognitive engagement, relying on intuition or heuristic judgments, which may lead to a decline in decision quality; whereas reasoning models maintain a relatively stable level of deliberation regardless of task difficulty, continuing to conduct in-depth analysis and reasoning.

Human deliberation allocation is subject to multiple constraints: awareness of cognitive costs (evaluating the energy consumption of thinking, reducing investment when it exceeds benefits), limited attention resources (working memory capacity limits high-intensity cognitive activities), and fluctuating motivation (fatigue and boredom affect persistence); reasoning models have no biological limitations, no concept of fatigue or boredom, and do not reduce investment due to task difficulty.

1. In scenarios requiring continuous high-quality reasoning (scientific research, complex data analysis, medical diagnosis assistance), reasoning models can complement human cognitive abilities; 2. Optimize human-AI collaboration task allocation: assign continuous deep thinking tasks to AI, while humans focus on creative insights and intuitive judgments; 3. Help understand human cognitive limitations and develop targeted cognitive training programs.

Limitations: Only focuses on the "quantity" rather than the "quality" of deliberation (how much is invested rather than its effectiveness); the results need to be verified when extended to complex real-world decision-making scenarios. Future directions: Explore how to maintain a high level of deliberation while improving reasoning quality and relevance.

This study reveals profound cognitive science issues. Understanding the essential differences and complementary advantages between humans and AI is key to human-AI collaboration. When humans need to "save brainpower", reasoning models can become tireless thinking partners—this is the promising future of human-AI collaboration.