Fudan Proves LLMs Fake Curiosity but CoQ Boosts Qwen2.5 Reasoning by 23 Points
Researchers adapt a psychology questionnaire and behavioral experiments to probe LLM curiosity, then turn the findings into a prompting trick that lifts reasoning accuracy by up to 23 points.
Takeaways
- First systematic curiosity evaluation for LLMs, adapting the psychology-validated 5DCR questionnaire plus three behavioral games.
- LLMs show strong information seeking but weak thrill seeking, an asymmetry the authors attribute to safety training.
- Logit-lens analysis shows curiosity signals only appear in the last few layers, indicating no intrinsic trait.
- Persona prompts easily reshape curiosity scores, confirming the behavior is context-dependent surface fitting.
- Curious Chain-of-Questioning prompting lifts Qwen2.5-7B by +23.7 points on DetectBench and +12.4 on Math.
- Code, data, and training configs released at Yukijudaii1352/CuriosityEval.
A team from Fudan University, Cornell, and Huawei just ran what may be the most thorough psychology experiment ever conducted on language models. They handed GPT-4o, Gemini 2.5, DeepSeek, Llama, Gemma, InternLM, and the Qwen family the same curiosity questionnaire clinical psychologists give to humans, then dropped them into behavioral games designed for children and adults. The verdict: LLMs act curious, but they don't seem to be curious in any deep sense, and that distinction matters for reasoning performance.
The paper, titled "Why Did Apple Fall," is motivated by a growing trend in reinforcement learning research. Recent training paradigms like GRPO and DAPO have spawned intrinsic reward signals inspired by human curiosity, such as i-MENTOR and CDE, that encourage action trajectories that are more uncertain or less frequently explored. Nobody has actually checked whether the behavior being reinforced resembles curiosity at all, or whether the gains just come from cranking up entropy on the loss function.
Building a psychology lab for language models
The framework starts from the Five-Dimensional Curiosity Scale Revised (5DCR), the most recent and authoritative scale in psychological research for measuring curiosity. The authors collapse its six subscales into three overarching dimensions worth defining:
- Information Seeking: joyous exploration of new ideas plus frustration when a knowledge gap can't be closed.
- Thrill Seeking: willingness to take risks or accept uncertainty for novel experiences.
- Social Curiosity: interest in other people's thoughts, behaviors, and relationships.

Because LLMs are notorious for saying whatever sounds agreeable on a self-report, the team doesn't stop at the questionnaire. They note that questionnaire responses can be influenced by hallucinated personas, item phrasing, and response scales, so they built three behavioral tasks lifted straight from psychology literature: