Banishing LLM Hallucinations Requires Rethinking Generalization

2024-06-25

arXiv.org

AI Summary

This paper challenges the conventional view that LLM hallucinations arise from a trade-off between creativity and factual accuracy, arguing instead that hallucination is fundamentally a failure of generalization under the training loss threshold. Method: We propose MoME (Mixture of Memory Experts), a hybrid architecture leveraging massive, retrievable expert memories; it enables dynamic sparse retrieval to realize “memory-as-fact” and “retrieval-as-reasoning.” Contribution/Results: We provide the first theoretical and empirical refutation of dominant attribution theories, establishing a causal link between hallucination and training loss thresholds. Building upon this insight, we develop Lamini-1—the first hallucination-mitigated LLM—which significantly reduces hallucination rates across multiple factuality benchmarks. Our results validate a new paradigm: replacing internal parametric knowledge with scalable, external memory systems for robust factual grounding.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Despite their powerful chat, coding, and reasoning abilities, Large Language Models (LLMs) frequently hallucinate. Conventional wisdom suggests that hallucinations are a consequence of a balance between creativity and factuality, which can be mitigated, but not eliminated, by grounding the LLM in external knowledge sources. Through extensive systematic experiments, we show that these traditional approaches fail to explain why LLMs hallucinate in practice. Specifically, we show that LLMs augmented with a massive Mixture of Memory Experts (MoME) can easily memorize large datasets of random numbers. We corroborate these experimental findings with a theoretical construction showing that simple neural networks trained to predict the next token hallucinate when the training loss is above a threshold as it usually does in practice when training on internet scale data. We interpret our findings by comparing against traditional retrieval methods for mitigating hallucinations. We use our findings to design a first generation model for removing hallucinations -- Lamini-1 -- that stores facts in a massive mixture of millions of memory experts that are retrieved dynamically.

Problem

• LLMs hallucinate despite external knowledge grounding
• Memorizing random data reveals fundamental generalization flaws
• Training loss threshold causes neural networks to hallucinate

Innovation

• Mixture of Memory Experts (MoME) architecture
• Dynamically retrieved millions of memory experts
• Stores facts in massive specialized memory system