Confabulation: The Surprising Value of Large Language Model Hallucinations

2024/06/06 by Peiqi Sui, Eamon Duede, Sui, Peiqi +5 · 8 voices · 19 citations
Computer Science · Mathematics · Neuroscience · Psychology · #Cognition #Cognitive psychology #Confabulation (neural networks) #Mathematics #Natural Language Processing Techniques #Neuroscience #Psychology #Semantic Web and Ontologies #Statistics #Topic Modeling #Value (mathematics)

paper · pdf · doi:10.48550/arxiv.2406.04175

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2024/06/06 · openalex created_date 2024/06/08 · openalex updated_date 2026/07/28

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.

Cited by

• Why Slop Matters
• Critical Confabulation: Can LLMs Hallucinate for Social Good?
• Technological folie à deux: Feedback Loops Between AI Chatbots and Mental Illness
• KRISTEVA: Close Reading as a Novel Task for Benchmarking Interpretive Reasoning
• Sycophantic AI decreases prosocial intentions and promotes dependence
• Provocations from the Humanities for Generative AI Research
• PENDULUM: A Benchmark for Assessing Sycophancy in Multimodal Large Language Models
• Simulating Subjects: The Promise and Peril of Artificial Intelligence Stand-Ins for Social Agents and Interactions
• Explainability of Large Language Models: Opportunities and Challenges toward Generating Trustworthy Explanations
• ReFACT: A Benchmark for Scientific Confabulation Detection with Positional Error Annotations
• Library Hallucinations in LLMs: Risk Analysis Grounded in Developer Queries
• MetaRAG: Metamorphic Testing for Hallucination Detection in RAG Systems
• LongRecall: A Structured Approach for Robust Recall Evaluation in Long-Form Text
• Can LLMs Detect Their Confabulations? Estimating Reliability in Uncertainty-Aware Language Models
• Personalized Constitutionally-Aligned Agentic Superego: Secure AI Behavior Aligned to Diverse Human Values
• Neither Stochastic Parroting nor AGI: LLMs Solve Tasks through Context-Directed Extrapolation from Training Data Priors
• AI Virtue: What is "Good" Knowledge in the Age of Artificial Intelligence?
• MemArena: An Ego-Centric Benchmark for On-Device Agentic Personal Memory Assistants at Scale
• Synthetic Fluency: Hallucinations, Confabulations, and the Creation of Irish Words in LLM-Generated Translations

Discussions

• A team centered at McGill and Harvard argues that LLM "hallucination" is better understood as confabulation, and resembles human habits of using narrative for sense-making. #machinelearning 🤖 arxiv.o [bsky, 36 points, 3 comments]
• Large language model (LLM) hallucinations are "a potential resource instead of a categorically negative pitfall" arxiv.org/pdf/2406.04175 hmm [bsky, 7 points, 2 comments]
• Confabulation: The Surprising Value of Large Language Model Hallucinations https://arxiv.org/abs/2406.04175v1 #AI #hallucination [bsky, 6 points, 1 comments]
• Totally selfish plug for our "Confabulation" main conference paper at #ACL2024 on exactly this issue: arxiv.org/abs/2406.04175 [bsky, 4 points, 1 comments]
• *Hallucinations: not a bug, but a feature arxiv.org/pdf/2406.04175 [bsky, 2 points, 0 comments]
• https://arxiv.org/abs/2406.04175 [bsky, 0 points, 0 comments]
• Some notes about this. The lead author is a PhD student who came up with the term "confabulations" as an alternative to "hallucinations," in this paper that says: "Most hallucinations are wrong, BUT S [bsky, 0 points, 1 comments]
• The Surprising Value of Large Language Model Hallucinations: the tendency for LLMs to confabulate may be intimately associated with a positive capacity for coherent narrative-text generation. arxiv.or [bsky, 0 points, 0 comments]

Related

• How Large Language Models are Designed to Hallucinate
• Large Language Models Hallucination: A Comprehensive Survey
• The Hallucinations Leaderboard -- An Open Effort to Measure Hallucinations in Large Language Models
• 🧜Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models
• A comprehensive taxonomy of hallucinations in Large Language Models