A Fictional Q&A Dataset for Studying Memorization and Knowledge Acquisition

2025-06-05

AI Summary

This work investigates the dual mechanisms of factual (semantic) memory versus sequential (verbatim) memory in language model training, and their decoupling. To this end, we introduce the first synthetic fictional Q&A dataset specifically designed to isolate these memory types: using a controllable fictional event generation paradigm, we eliminate interference from real-world knowledge, thereby enabling systematic differentiation between factual knowledge acquisition and literal text repetition. Methodologically, our approach integrates synthetic data construction, controlled text modeling, training trajectory tracking, and empirical memory-behavior evaluation. Experiments demonstrate that the dataset effectively disentangles the two memory pathways; further, they reveal substantial memory biases in current models—such as overreliance on superficial lexical cues—and critical generalization bottlenecks in transferring learned fictional facts. This study establishes a novel, reproducible paradigm and benchmark for probing the memory architecture of large language models.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Computational Creativity

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

When language models are trained on textual data, they acquire both knowledge about the structure of language as well as knowledge of facts about the world. At inference time, their knowledge of facts can be leveraged to solve interesting problems and perform useful knowledge work for users. It is well known that language models can verbatim memorize long sequences from their training data. However, it is much less well understood how language models memorize facts seen during training. In this work, we propose a new dataset to specifically empower researchers to study the dual processes of fact memorization and verbatim sequence memorization. The dataset consists of synthetically-generated, webtext-like documents about fictional events, as well as question-answer pairs about the events. We conduct training experiments showing how synthetic data about fictional events can be effective in teasing apart different forms of memorization. We also document the challenges in effectively building realistic, fictional synthetic data.

Problem

• Studying how language models memorize facts during training
• Differentiating fact memorization from verbatim sequence memorization
• Creating realistic synthetic data for memorization research

Innovation

• Synthetic Q&A dataset for memorization study
• Fictional events to separate memorization types
• Webtext-like documents for realistic simulation