Does Less Hallucination Mean Less Creativity? An Empirical Investigation in LLMs

Technical Summary: "Does Less Hallucination Mean Less Creativity? An Empirical Investigation in LLMs"

Problem Statement
Large Language Models (LLMs) are increasingly utilized for natural language understanding and reasoning, yet they suffer from "hallucination"—the generation of factually incorrect content. While significant research has focused on mitigating hallucinations to ensure factual reliability, particularly in high-stakes domains like AI-assisted scientific discovery, the impact of these mitigation strategies on model creativity remains unexplored. This is a critical gap because scientific discovery relies on a dual requirement: factual accuracy and the creative generation of hypotheses, which often involves making unconventional connections that may superficially resemble hallucinations. The paper investigates whether reducing hallucinations inadvertently suppresses the divergent thinking necessary for creative ideation.

Methodology
The authors conducted an empirical investigation evaluating the effects of three distinct hallucination-reduction techniques on creativity across multiple model families and scales:

1. Techniques Evaluated:

   • Chain of Verification (CoVe): A multi-stage reasoning process involving drafting, generating verification questions, answering them, and synthesizing a refined response.
   • Decoding by Contrasting Layers (DoLa): A decoding method that contrasts predictions from the final layer against earlier layers (specifically the one with the highest Jensen-Shannon divergence) to suppress less reliable tokens.
   • Retrieval-Augmented Generation (RAG): A method grounding generation in external knowledge retrieved from a corpus (specifically coding tutorials and documentation in this study).

2. Benchmarks and Metrics:

   • NeoCoder: A programming benchmark evaluating creativity under progressively constrained rule-based scenarios (analogous to scientific laws). Metrics included Convergent Creativity (correctness under constraints) and Divergent Creativity (novelty of techniques beyond human patterns).
   • CS4: A story generation benchmark evaluating open-ended creative writing under cumulative constraints. Metrics included Constraint Satisfaction, Coherence, Diversity (DIST-N), and a composite Creativity score (QUC).

3. Experimental Scope:

   • Models: LLaMA (1B, 8B, 70B), Qwen-Coder (7B), and Mistral (7B).
   • Analysis: The study measured percentage changes in creativity metrics relative to a baseline (no hallucination reduction). Additionally, the authors employed linear probes to identify which transformer layers correlate with creative output, leading to a modified decoding strategy.

Key Results
The investigation revealed that hallucination-reduction methods have divergent and opposing effects on divergent creativity, while largely preserving convergent creativity (the ability to solve problems correctly).

• CoVe Enhances Divergent Creativity: Contrary to the initial hypothesis that fact-checking might stifle creativity, CoVe consistently improved divergent thinking across most models and datasets. For instance, LLaMA 1B showed up to a 12.5% improvement on NeoCoder. The authors attribute this to the questioning mechanism encouraging broader exploration of the solution space and reducing "tunnel vision."
• DoLa Suppresses Divergent Creativity: DoLa consistently reduced divergent creativity (e.g., LLaMA 1B dropped ~8% on CS4). The authors hypothesize this occurs because DoLa contrasts early and late layers; since early layers encode exploratory and divergent representations, subtracting them inadvertently removes features necessary for creative generation.
• RAG Shows Minimal Impact: RAG demonstrated a neutral effect on divergent creativity, with fluctuations near zero. The authors suggest this is due to a semantic mismatch between the retrieval corpus (technical tutorials) and the task (narrative problem descriptions), resulting in retrieved documents that lacked specific algorithmic insights or introduced noise.
• Convergent Creativity is Unaffected: Across all methods and models, the ability to generate factually correct and constraint-satisfying solutions remained largely unchanged.

Further Investigation: Layer Modulation
To address the suppression caused by DoLa, the authors used linear probes to identify "creativity-correlated" layers. They found that early transformer layers exhibited stronger correlations with divergent creativity than later layers. By reversing the DoLa logic—amplifying activations in creativity-correlated layers while suppressing anti-correlated ones—they successfully boosted divergent creativity without compromising convergent performance. This suggests that divergent and convergent creativity can be decoupled and independently modulated.

Significance and Claims
The paper claims that the relationship between hallucination reduction and creativity is not uniform; it depends heavily on the specific technique employed.

• For AI4Science: The findings provide guidance for selecting hallucination-reduction methods in scientific applications. Techniques like CoVe may be preferable for hypothesis generation as they enhance creative exploration without sacrificing factual convergence, whereas DoLa may be detrimental to the imaginative phase of discovery.
• Inherent Characteristics: The consistency of these trends across different model families (LLaMA, Qwen, Mistral) and scales (1B–70B) suggests that the interaction between factual control and creative reasoning is an inherent characteristic of LLMs, rather than an artifact of architecture or size.
• Future Direction: The authors emphasize that while programming and story generation serve as proxies, future work must develop frameworks specifically for scientific hypothesis generation to validate these findings in authentic discovery contexts. They also note the need for ablation studies to isolate the specific mechanisms by which questioning enhances creativity and for improved retrieval strategies to maximize RAG's potential.

In conclusion, the paper argues that reducing hallucinations does not necessarily equate to reducing creativity; rather, the choice of mitigation strategy determines whether a model becomes a precise but limited tool or a balanced collaborator capable of both factual accuracy and creative ideation.