Pre-training Limited Memory Language Models with Internal and External Knowledge

2025/05/21 by Zhao, Linxi, Sofian Zalouk, Christian K. Belardi +14 · 1 voice · 3 citations
Computer Science · #Topic Modeling #Big Data and Digital Economy #Natural Language Processing Techniques

paper · pdf · doi:10.48550/arxiv.2505.15962

Abstract

Neural language models are black-boxes--both linguistic patterns and factual knowledge are distributed across billions of opaque parameters. This entangled encoding makes it difficult to reliably inspect, verify, or update specific facts. We introduce Limited Memory Language Models (LMLM), a new class of language models that externalizes factual knowledge to external database during pre-training rather than memorizing them. Our pre-training approach strategically masks externally retrieved factual values from the training loss, thereby teaching the model to perform targeted lookups rather than relying on memorization in model weights. Our experiments demonstrate that LMLMs achieve competitive performance compared to significantly larger LLMs on standard benchmarks, while offering the advantages of explicit, editable, and verifiable knowledge bases.

Citations

• Memorization Sinks: Isolating Memorization during LLM Training
• Parameters vs. Context: Fine-Grained Control of Knowledge Reliance in Language Models
• Measuring short-form factuality in large language models
• Advancing Large Language Model Attribution through Self-Improving
• ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability
• Do Unlearning Methods Remove Information from Language Model Weights?
• Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning
• Mechanistic Interpretability for AI Safety -- A Review
• Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws
• Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
• Understanding Emergent Abilities of Language Models from the Loss Perspective
• Blinded by Generated Contexts: How Language Models Merge Generated and Retrieved Contexts When Knowledge Conflicts?
• TOFU: A Task of Fictitious Unlearning for LLMs
• Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs
• Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection
• Physics of Language Models: Part 3.2, Knowledge Manipulation
• Physics of Language Models: Part 3.1, Knowledge Storage and Extraction
• Large Language Models
• Unifying Large Language Models and Knowledge Graphs: A Roadmap
• Direct Preference Optimization: Your Language Model is Secretly a Reward Model
• FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
• Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts
• Complex Logical Reasoning over Knowledge Graphs using Large Language Models
• A Note on Normalized Emergence Timing (in Pythia Language Model Evaluations)
• Learn to Memorize: Scalable Continual Learning in Semiparametric Models with Mixture-of-Neighbors Induction Memory
• Check Your Facts and Try Again: Improving Large Language Models with External Knowledge and Automated Feedback
• REPLUG: Retrieval-Augmented Black-Box Language Models
• When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories
• Knowledge-in-Context: Towards Knowledgeable Semi-Parametric Language Models
• Mass-Editing Memory in a Transformer
• Knowledge Unlearning for Mitigating Privacy Risks in Language Models
• Locating and Editing Factual Associations in GPT
• Improving language models by retrieving from trillions of tokens
• Fast Model Editing at Scale
• Entity-Based Knowledge Conflicts in Question Answering
• Finetuned Language Models Are Zero-Shot Learners
• LoRA: Low-Rank Adaptation of Large Language Models
• Distilling Knowledge from Reader to Retriever for Question Answering
• CoLAKE: Contextualized Language and Knowledge Embedding
• KILT: a Benchmark for Knowledge Intensive Language Tasks
• Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering
• Language Models are Few-Shot Learners
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• BERT-kNN: Adding a kNN Search Component to Pretrained Language Models for Better QA
• Entities as Experts: Sparse Memory Access with Entity Supervision
• Editable Neural Networks
• REALM: Retrieval-Augmented Language Model Pre-Training
• K-Adapter: Infusing Knowledge into Pre-Trained Models with Adapters
• PIQA: Reasoning about Physical Commonsense in Natural Language
• Generalization through Memorization: Nearest Neighbor Language Models
• Language Models as Knowledge Bases?
• HellaSwag: Can a Machine Really Finish Your Sentence?
• Know What You Don't Know: Unanswerable Questions for SQuAD
• Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge
• RAGGED: Towards Informed Design of Scalable and Stable RAG Systems
• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
• CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge

Cited by

• Rotation Control Unlearning: Quantifying and Controlling Continuous Unlearning for LLM with The Cognitive Rotation Space
• Pretraining with hierarchical memories: separating long-tail and common knowledge
• Co-LMLM: Continuous-Query Limited Memory Language Models

Discussions

• Pre-Training Large Memory Language Models with Internal and External Knowledge [hn, 4 points, 0 comments]

Related

• How Do Large Language Models Acquire Factual Knowledge During Pretraining?
• Memorisation versus Generalisation in Pre-trained Language Models
• Towards a Holistic Evaluation of LLMs on Factual Knowledge Recall
• Augmenting LLMs with Knowledge: A survey on hallucination prevention
• What Matters in Memorizing and Recalling Facts? Multifaceted Benchmarks for Knowledge Probing in Language Models