Too Big to Think: Capacity, Memorization, and Generalization in Pre-Trained Transformers

2025-06-10

AI Summary

This work investigates how model capacity governs the intrinsic trade-off between memorization and generalization in large language models (LLMs). We systematically pretrain multiple families of capacity-constrained Transformer models from scratch within a unified framework and evaluate them on synthetic character-level arithmetic extrapolation (generalization) and factual recall (memorization) tasks. Our study is the first to empirically uncover a capacity-driven memorization–generalization trade-off: smaller models exhibit strong generalization but weak memorization, larger models show the opposite, and medium-capacity models also exhibit pronounced memorization bias. Crucially, joint training fails to recover extrapolation capability, revealing an inherent memorization bias in standard pretraining paradigms. By leveraging synthetic data, decoupled task design, and exhaustive capacity scanning across the full spectrum, our work provides reproducible, causal evidence for understanding the fundamental learning dynamics of LLMs.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

The relationship between memorization and generalization in large language models (LLMs) remains an open area of research, with growing evidence that the two are deeply intertwined. In this work, we investigate this relationship by pre-training a series of capacity-limited Transformer models from scratch on two synthetic character-level tasks designed to separately probe generalization (via arithmetic extrapolation) and memorization (via factual recall). We observe a consistent trade-off: small models extrapolate to unseen arithmetic cases but fail to memorize facts, while larger models memorize but fail to extrapolate. An intermediate-capacity model exhibits a similar shift toward memorization. When trained on both tasks jointly, no model (regardless of size) succeeds at extrapolation. These findings suggest that pre-training may intrinsically favor one learning mode over the other. By isolating these dynamics in a controlled setting, our study offers insight into how model capacity shapes learning behavior and offers broader implications for the design and deployment of small language models.

Problem

• Investigates memorization-generalization trade-off in Transformers
• Tests capacity impact on arithmetic extrapolation and factual recall
• Explores pre-training bias toward memorization over generalization

Innovation

• Pre-trained capacity-limited Transformers on synthetic tasks
• Isolated generalization and memorization via arithmetic extrapolation
• Analyzed model capacity impact on learning behavior