TLDR: A new benchmark called “Rationality Check!” has been developed to evaluate the rationality of Large Language Models (LLMs) across six domains: psychology, cognitive science, decision-making, economics, game theory, and collective rationality. The study found that LLMs generally exhibit high rationality, especially larger models, and excel in areas like self-reflection, emotional reappraisal, and collective cooperation. However, they still lag behind humans in core logical reasoning tasks and struggle with complex strategic games. The benchmark highlights both convergences and divergences between LLM and human rationality, offering a critical tool for improving AI alignment and application suitability.
Large Language Models (LLMs) have made incredible strides in mimicking human-like abilities, from understanding and generating natural language to tackling complex reasoning and planning tasks. As these AI systems become more integrated into our lives, simulating human activities and serving as AI assistants, a crucial question arises: how rational are they? A recent research paper, titled “Rationality Check! Benchmarking the Rationality of Large Language Models,” introduces the first comprehensive benchmark designed to answer this very question.
The concept of rationality is fundamental to understanding human behavior, encompassing both theoretical rationality (how we think) and practical rationality (how we act). This new benchmark aims to evaluate the overall rationality of LLMs across a broad spectrum of domains and models, providing a foundational tool for both developers and users of these powerful AI systems. The researchers have also made their code and dataset publicly available for further exploration. You can find more details in the full research paper: Rationality Check! Benchmarking the Rationality of Large Language Models.
The benchmark organizes rationality into six distinct domains: psychology, cognitive science, decision-making, economics, game theory, and collective rationality. Each domain includes various aspects, categorized into theoretical and practical rationality. Theoretical rationality focuses on the reasonableness of internal beliefs, while practical rationality emphasizes appropriate decision-making in real-world contexts.
Also Read:
- Fables Expose Flaws in Large Language Models’ Moral Understanding
- Nazonazo: Japanese Riddles Uncover LLM’s Insight Problem
Exploring Rationality Across Domains
In the **psychology domain**, LLMs generally show higher self-reflection and emotional reappraisal abilities than humans. This means they are better at introspection and reframing emotional situations. However, advanced models surprisingly struggle with emotional suppression, and some models are less motivated for complex cognitive tasks compared to humans.
The **cognitive science domain** reveals that advanced LLMs, such as GPT-4o and DeepSeek v2.5, excel in rational thinking, scientific reasoning, abstract reasoning, and critical thinking. They also demonstrate lower susceptibility to certain cognitive biases like regret aversion. Interestingly, humans still outperform LLMs in fundamental logical reasoning tasks like inductive, deductive, and causal reasoning, suggesting areas for future LLM improvement. While LLMs are generally better at avoiding cognitive biases, some advanced models still struggle with biases like illusion of control and belief bias, indicating challenges in applying theoretical rationality to practical scenarios.
For **decision-making**, LLMs perceive themselves as more rational than humans. Newer models tend to be avoidant, dependent, and spontaneous decision-makers, though less intuitive than humans (except for Qwen models). They show significantly lower susceptibility to most decision-making biases, with exceptions in the Better-than-Average Effect and Causal Base-Rate Neglect.
In the **economics domain**, LLMs are found to be overly concerned with moral and health safety risks, often more so than humans. However, their actual risk-taking behavior in action doesn’t always align with this high level of concern, sometimes showing less caution than humans.
The **game theory domain** presents a different picture, with most LLMs exhibiting lower rationality than humans. Even top-performing models like GPT-4 struggle in strategic games that require anticipating others’ actions. Only one model, text-bison, surpassed human rationality in this domain. This suggests that game theory scenarios are particularly challenging for LLMs. The study also found that larger models tend to be more rational, and fine-tuning processes can positively impact an LLM’s rationality in game theory.
Finally, in the **collective rationality domain**, LLMs generally outperform humans in cooperation and coordination games, likely due to their consistent and stable strategies. Larger models also demonstrate higher rationality in “wisdom of crowds” experiments, with models like GPT-4 and GPT-3.5 significantly outperforming humans in general knowledge and math problems. LLMs tend to answer general knowledge questions correctly regardless of agent profiles, unlike humans whose answers vary.
In conclusion, this benchmark provides valuable insights into the multifaceted rationality of LLMs. While they show impressive capabilities in many areas, particularly in theoretical reasoning and collective tasks, there are still notable divergences from human rationality, especially in complex strategic interactions and certain practical applications. This research is crucial for guiding the development of safer and more responsible AI applications, helping developers train better LLMs for human-AI alignment and assisting users in selecting appropriate LLMs for specific tasks.
