Heaven-Sent or Hell-Bent? Benchmarking the Intelligence and Defectiveness of LLM Hallucinations

TL;DR

The paper reframes LLM hallucinations as a spectrum that can fuel scientific creativity by separating Intelligent Hallucinations (IH) from Defective Hallucinations (DH). It introduces HIC-Bench, a cross-domain benchmark combining TTCT-inspired creativity metrics with a multi-dimensional IH/DH framework and a Dynamic Hallucination Prompt (DHP) to steer outputs toward valuable novelty. Using the Cross-Domain Innovation Dataset (CDID) and Cross-Domain Knowledge Base (CDKB), the authors evaluate six diverse LLMs across ten domains, reporting IH, DH, and an Intelligent-Fidelity Score (IFS). Results reveal a nonlinear relationship between IH and DH and show that DHP, especially when paired with moderated constraints, can boost IH while reducing DH, highlighting the potential to harness hallucinations as a driver of innovation rather than merely suppressing them.

Abstract

Hallucinations in large language models (LLMs) are commonly regarded as errors to be minimized. However, recent perspectives suggest that some hallucinations may encode creative or epistemically valuable content, a dimension that remains underquantified in current literature. Existing hallucination detection methods primarily focus on factual consistency, struggling to handle heterogeneous scientific tasks and balance creativity with accuracy. To address these challenges, we propose HIC-Bench, a novel evaluation framework that categorizes hallucinations into Intelligent Hallucinations (IH) and Defective Hallucinations (DH), enabling systematic investigation of their interplay in LLM creativity. HIC-Bench features three core characteristics: (1) Structured IH/DH Assessment. using a multi-dimensional metric matrix integrating Torrance Tests of Creative Thinking (TTCT) metrics (Originality, Feasibility, Value) with hallucination-specific dimensions (scientific plausibility, factual deviation); (2) Cross-Domain Applicability. spanning ten scientific domains with open-ended innovation tasks; and (3) Dynamic Prompt Optimization. leveraging the Dynamic Hallucination Prompt (DHP) to guide models toward creative and reliable outputs. The evaluation process employs multiple LLM judges, averaging scores to mitigate bias, with human annotators verifying IH/DH classifications. Experimental results reveal a nonlinear relationship between IH and DH, demonstrating that creativity and correctness can be jointly optimized. These insights position IH as a catalyst for creativity and reveal the ability of LLM hallucinations to drive scientific innovation.Additionally, the HIC-Bench offers a valuable platform for advancing research into the creative intelligence of LLM hallucinations.

Figures (5)

• Figure 1: Example Instances of Hallucination Types, Evaluation, and Review in HIC-Bench
• Figure 2: Design of the HIC-Bench Framework: We first construct a cross-disciplinary dataset by integrating Wikipedia principles with real-world problems. We then evaluate the model’s responses across various generation strategies, analyzing its creativity and hallucinations through automated evaluation. Finally, human reviewers perform token-level analysis on the model’s evaluation outputs to enhance their reliability.“Or” for Originality, “Fe” for Feasibility, “Va” for Value, “Flu” for Fluency, and “Fle” for Flexibility.
• Figure 3: Comparative Analysis of Originality and IH Across Scientific Domains
• Figure 4: Comparative analysis of model indicators across scientific fields under SCP
• Figure 5: Comparative analysis of model indicators across scientific fields under SCP