Productive LLM Hallucinations: Conditions, Mechanisms, and Benefits

Feng He, Zhenting Wang, Qifan Wang, Qiang Guan, Dongfang Liu, Ruixiang Tang, Qiankun Li

Keywords: Large Language Models; Hallucination; Productive Hallucinations; Reasoning Dynamics

TL;DR: We introduce HIVE, a framework showing that moderate hallucinations in LLMs can broaden semantics, stabilize reasoning, and improve accuracy across diverse tasks, reframing them as a controllable cognitive resource.

Abstract: Hallucinations in large language models (LLMs) are typically regarded as harmful errors to be suppressed. We revisit this assumption and ask whether, and under what conditions, hallucinations can instead be beneficial. To address this question, we introduce $\textbf{HIVE}$ ($\textbf{H}$allucination $\textbf{I}$nference and $\textbf{V}$erification $\textbf{E}$ngine), a task-agnostic framework that systematically evaluates the impact of hallucinated semantics across diverse tasks and models. By unifying generation, discrimination, and downstream evaluation, HIVE enables controlled comparative assessments of how hallucinations alter overall model performance. Extensive experiments on nine datasets and ten models show that hallucinations can yield substantial improvements up to $\textbf{+17.2}$ \% in accuracy especially in open-ended domains such as reasoning, biomedical, and vision language tasks. Stronger models consistently harness hallucinations, while weaker ones are more volatile. Mechanistic analyses show that hallucinations broaden semantic coverage, stabilize reasoning trajectories, and follow an inverted-U profile where moderate strength maximizes benefits across diverse tasks. These findings reframe hallucination from a defect to a controllable cognitive resource, suggesting opportunities for evaluating and training LLMs not merely to avoid hallucinations, but to exploit them constructively.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 468