Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Confabulation: The Surprising Value of Large Language Model Hallucinations

Peiqi Sui, Eamon Duede, Sophie Wu, Richard Jean So

TL;DR

It is argued and empirically demonstrate that measurable semantic characteristics of LLM confabulations mirror a human propensity to utilize increased narrativity as a cognitive resource for sense-making and communication, and suggests that the tendency for LLMs to confabulate may be intimately associated with a positive capacity for coherent narrative-text generation.

Abstract

This paper presents a systematic defense of large language model (LLM) hallucinations or 'confabulations' as a potential resource instead of a categorically negative pitfall. The standard view is that confabulations are inherently problematic and AI research should eliminate this flaw. In this paper, we argue and empirically demonstrate that measurable semantic characteristics of LLM confabulations mirror a human propensity to utilize increased narrativity as a cognitive resource for sense-making and communication. In other words, it has potential value. Specifically, we analyze popular hallucination benchmarks and reveal that hallucinated outputs display increased levels of narrativity and semantic coherence relative to veridical outputs. This finding reveals a tension in our usually dismissive understandings of confabulation. It suggests, counter-intuitively, that the tendency for LLMs to confabulate may be intimately associated with a positive capacity for coherent narrative-text generation.

Confabulation: The Surprising Value of Large Language Model Hallucinations

TL;DR

It is argued and empirically demonstrate that measurable semantic characteristics of LLM confabulations mirror a human propensity to utilize increased narrativity as a cognitive resource for sense-making and communication, and suggests that the tendency for LLMs to confabulate may be intimately associated with a positive capacity for coherent narrative-text generation.

Abstract

This paper presents a systematic defense of large language model (LLM) hallucinations or 'confabulations' as a potential resource instead of a categorically negative pitfall. The standard view is that confabulations are inherently problematic and AI research should eliminate this flaw. In this paper, we argue and empirically demonstrate that measurable semantic characteristics of LLM confabulations mirror a human propensity to utilize increased narrativity as a cognitive resource for sense-making and communication. In other words, it has potential value. Specifically, we analyze popular hallucination benchmarks and reveal that hallucinated outputs display increased levels of narrativity and semantic coherence relative to veridical outputs. This finding reveals a tension in our usually dismissive understandings of confabulation. It suggests, counter-intuitively, that the tendency for LLMs to confabulate may be intimately associated with a positive capacity for coherent narrative-text generation.
Paper Structure (14 sections, 1 figure, 4 tables)

This paper contains 14 sections, 1 figure, 4 tables.

Table of Contents

  1. Background
  2. Related Work
  3. Confabulation vs Hallucination
  4. Towards a Narrative-Centered Definition of Confabulation
  5. Data, Methods, and Results
  6. Empirical Results for Higher Narrativity in Hallucinations
  7. In defense of confabulation
  8. Empirical Support for Association of Narrativity and Coherence in Confabulated Texts
  9. Narrative, Discourse, and Coherence
  10. Narratives Help Us Articulate and Understand Complex Arguments
  11. Narratives Maintain the Consistency of Our Own Internal World Models
  12. Narratives Enable Patients to Negotiate the Coherence of Their Experiences
  13. Limitations and Directions for Future Research
  14. Acknowledgement

Figures (1)

  • Figure 1: The left panel illustrates distribution for narrative score of hallucinated outputs (blue) and the edited version of the output (gray) in the FaithDial dataset. The hallucinated texts are, in general more narrative rich than those that are edited to resolve inaccuracies. The right panel illustrates distribution for non-hallucinated texts from the FaithDial dataset.