The mosaic memory of large language models.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Large Language Models (LLMs) are increasingly prevalent.
• Understanding LLM memorization of training data is critical.
• Current assumptions link memorization solely to data repetition.

Purpose of the Study:

• To investigate the mechanisms of memorization in LLMs.
• To introduce and define the concept of 'mosaic memory'.
• To analyze the implications of LLM memorization for data privacy and model evaluation.

Main Methods:

• Investigated memorization in major LLMs.
• Quantified the contribution of fuzzy duplicates to memorization.
• Analyzed the syntactic versus semantic nature of memorization.
• Examined the prevalence of fuzzy duplicates in real-world data.

Main Results:

• LLMs exhibit 'mosaic memory,' assembling information from similar, not just identical, sequences.
• Fuzzy duplicates contribute significantly to memorization, comparable to exact duplicates.
• Memorization is predominantly syntactic, not semantic, despite LLMs' reasoning abilities.
• Ubiquitous fuzzy duplicates exist in real-world data, evading deduplication.

Conclusions:

• LLM memorization is a complex mosaic process, not solely based on exact repetition.
• Mosaic memory has significant implications for data privacy, confidentiality, and model utility.
• Current deduplication techniques are insufficient to address fuzzy duplicates in training data.
• Further research is needed to understand and mitigate the risks associated with mosaic memory.