The Dance of Hallucination and Creativity in LLMs' Decoding Layers via the Lens of Question Answering

ACL ARR 2025 May Submission6722 Authors

Abstract: Large language models (LLMs) are known to hallucinate, a phenomenon often linked to creativity. Built upon prior research that focuses on theoretical or qualitative analyses, our work uses a quantitative approach to systematically examine the relationship between hallucination and creativity in LLMs. Given the complex nature of creativity, we take the inspiration from philosophy and propose a creativity definition tailored to LLMs in Question Answering (QA) tasks. Further, we introduce an evaluation framework, HCL, to examine the relationship between Hallucination and Creativity across different Layers of LLMs during decoding. Our empirical analysis reveals a tradeoff between hallucination and creativity that is consistent across layer depth, model type, and model size. Notably, across different model architectures, we identify a specific layer at each model size that optimally balances this tradeoff. The optimal layer tends to appear in the early layers of larger models, and the confidence of the model is significantly higher at this layer. These findings provide a quantitative perspective that offers new insights into the interplay between LLM creativity and hallucination.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: hierarchical & concept explanations

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: English

Submission Number: 6722