TLDR: A new research paper evaluates curiosity in Large Language Models (LLMs) using human-inspired questionnaires and behavioral experiments. It finds that LLMs exhibit a stronger desire for knowledge than humans but are more conservative in uncertain environments. The study also demonstrates that curiosity-driven questioning significantly enhances LLMs’ reasoning and active learning capabilities, suggesting a novel approach for advancing AI intelligence.
Curiosity, a fundamental human trait, has driven countless discoveries throughout history, from Isaac Newton’s observation of a falling apple to scientific breakthroughs. It’s the intrinsic drive to seek information, ask questions, and explore new environments that fuels human learning and knowledge acquisition. In recent years, Large Language Models (LLMs) like ChatGPT have demonstrated impressive capabilities in understanding language and world knowledge. However, unlike humans who actively question and explore, LLMs primarily learn through predicting the next token in vast datasets, which might limit their potential for deeper, more generalized learning.
A new research paper, “Why Did Apple Fall To The Ground: Evaluating Curiosity In Large Language Model” by Haoyu Wang, Sihang Jiang, Yuyan Chen, Yitong Wang, and Yanghua Xiao from Fudan University, delves into whether these advanced AI models possess a human-like capacity for curiosity-driven learning. The study aims to systematically evaluate LLM curiosity, addressing a crucial gap in understanding their intrinsic motivations and potential for autonomous exploration.
A Comprehensive Framework for AI Curiosity
To assess LLM curiosity, the researchers developed a comprehensive evaluation framework inspired by the Five-Dimensional Curiosity Scale Revised (5DCR), a well-established human curiosity assessment questionnaire. This framework categorizes curiosity into three main dimensions: Information Seeking, Thrill Seeking, and Social Curiosity, further broken down into six sub-dimensions.
The evaluation involved two primary approaches:
- Questionnaire Study: LLMs were prompted to self-assess their curiosity levels using the 5DCR questionnaire, similar to how humans would respond.
- Behavioral Experiments: To overcome the potential unreliability of self-reports (due to factors like persona hallucination), the study adapted three psychology behavioral experiments for LLMs:
- Information Seeking: A “missing-letter game” where LLMs completed incomplete words and chose whether to view the answer, indicating their desire for unknown information.
- Thrill Seeking: An “underwater game” where LLMs selected between windows offering different levels of uncertainty about what fish they would see, measuring their willingness to take risks for novel experiences.
- Social Curiosity: A multi-round dialogue with a stranger (simulated by another LLM), where the frequency of questions asked by the LLM reflected its curiosity about others’ information in social scenarios.
Curiosity’s Role in AI Learning
Beyond just evaluating curiosity, the paper also explored its relationship with LLMs’ active learning abilities. The researchers compared two reasoning approaches: traditional Chain-of-Thought (CoT) and a novel Curious Chain-of-Questioning (CoQ). CoQ encourages LLMs to ask more auxiliary questions during their thought process, simulating a more inquisitive and exploratory learning style.
This investigation involved both prompt engineering for closed-source models and a combination of Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR) for open-source models, guiding them towards more curious thinking patterns.
Key Findings: How LLMs Compare to Humans
The study yielded several fascinating insights into the nature of AI curiosity:
- Stronger Information Seeking: LLMs generally exhibited a stronger desire for knowledge and information seeking than humans. In the missing-letter game, LLMs chose to peek at answers 70-80% of the time, compared to humans’ 37.8%.
- Conservative Thrill Seeking: In contrast, LLMs showed significantly lower thrill-seeking tendencies than humans. They preferred safer, more certain options in the underwater game, indicating a cautious stance towards uncertainty and risk.
- Comparable Social Curiosity: The social curiosity levels of LLMs were largely comparable to humans, with some models (like GPT and Gemini) even showing stronger proactive engagement in social questioning.
- Curiosity Enhances Learning: Perhaps the most significant finding was that curiosity-driven questioning (CoQ) substantially improved LLMs’ reasoning and active learning abilities. CoQ helped models avoid premature conclusions, prevented endless loops of reflection, and fostered “eureka” moments by identifying blind spots in conventional reasoning. This suggests that an inquisitive approach can broaden an LLM’s reasoning scope and lead to more insightful solutions.
In essence, while LLMs possess a robust thirst for knowledge, they tend to be more risk-averse than humans when faced with uncertainty. Crucially, cultivating this intrinsic curiosity through questioning can significantly boost their capacity for active learning and problem-solving, paving the way for new paradigms in AI development.
Also Read:
- AI’s Eye for Motives: How Language Models Learn to Discern Intent in Communication
- AI’s Moral Compass: Large Language Models Prioritize Emotion Over Cost in Fairness Decisions
Looking Ahead
This pioneering work provides experimental support for the future development of learning capabilities and innovative research in LLMs. While the study acknowledges limitations, such as the challenges of directly replicating human psychological experiments and the models’ sensitivity to prompts, it underscores the potential for LLMs to exhibit curiosity akin to humans, offering a promising direction for creating more autonomous and intelligent AI systems.
