Bring Your Own Knowledge: A Survey of Methods for LLM Knowledge Expansion

Mingyang Wang, Alisa Stoll, Lukas Lange, Heike Adel, Hinrich Schütze, Jannik Strötgen

Abstract

Adapting large language models (LLMs) to new and diverse knowledge is essential for their lasting effectiveness in real-world applications. This survey provides an overview of state-of-the-art methods for expanding the knowledge of LLMs, focusing on integrating various knowledge types, including factual information, domain expertise, language proficiency, and user preferences. We explore techniques, such as continual learning, model editing, and retrieval-based explicit adaptation, while discussing challenges like knowledge consistency and scalability. Designed as a guide for researchers and practitioners, this survey sheds light on opportunities for advancing LLMs as adaptable and robust knowledge systems.

Introduction: The Static Knowledge Problem

Large Language Models (LLMs) like GPT-4 or Llama are typically "frozen" at their training cutoff date. This static nature creates a significant hurdle: models struggle to incorporate new facts, adapt to specialized domains, or respect updated user preferences without undergoing a full, prohibitively expensive retraining process. While fine-tuning is a common alternative, it often suffers from "catastrophic forgetting," where new information overwrites previously learned skills. This survey explores "Knowledge Expansion"—a suite of techniques designed to keep models relevant and accurate across four key dimensions: factual truths, domain expertise, language proficiency, and user preferences.

A Taxonomy of Knowledge Expansion

The landscape of knowledge expansion can be categorized into three primary architectural approaches: Continual Learning, Model Editing, and Retrieval-based methods.

Taxonomy of Knowledge Expansion Methods

1. Continual Learning (CL)

Continual Learning involves incrementally updating a model's parameters as new data arrives. Beyond standard continual pretraining (CPT), advanced methods like parameter isolation (e.g., DEMix-DAPT) allocate specific subsets of neurons to new tasks or domains. This modular approach keeps the core model frozen, effectively mitigating the risk of forgetting.

2. Model Editing (ME)

Model Editing acts as "knowledge surgery," performing precise updates on specific weights rather than broad retraining. Locate-and-edit strategies, such as ROME and MEMIT, identify "knowledge neurons" within feed-forward layers. This allows for surgical corrections—like updating a world leader's name—without disturbing unrelated parametric knowledge.

3. Retrieval-based Methods

These methods bypass internal parameter changes by allowing the model to query external data during inference. Retrieval-Augmented Generation (RAG) is the most common implementation. More advanced techniques like kNN-LM interpolate the model's token distribution with results from a nearest-neighbor search in an external datastore, enabling real-time factual updates.

Comparative Analysis and Key Results

Each method presents distinct trade-offs regarding precision, scalability, and the type of knowledge it can effectively incorporate.

Comparison of Knowledge Expansion Methods

As shown in the comparison above, Continual Learning is the most robust for broad shifts like language expansion but lacks the precision required for individual fact updates. Conversely, Model Editing offers high precision for specific facts and preferences but fails to scale to large datasets or handle language expansion. Retrieval stands out as a versatile "black-box" solution that excels at factual and domain updates with low computational cost, though it currently struggles with expanding a model's core language or preference capabilities.

Key Insights and Challenges

  • The Precision-Scale Trade-off: There is a fundamental tension between broad adaptation and surgical precision. Model Editing struggles to scale to thousands of updates without degrading general reasoning, while CL lacks the control needed for single-fact corrections.
  • Implicit vs. Explicit Knowledge: CL and ME modify the model's internal (implicit) memory, whereas Retrieval relies on external (explicit) memory. For volatile data like stock prices, explicit retrieval is often the only viable path.
  • The "Reverse Curse": A notable finding in Model Editing is that updating a model to know "A is B" does not guarantee it understands "B is A." This suggests that knowledge in LLMs is often tied to specific token sequences rather than abstract, bidirectional concepts.
  • Knowledge Consistency: Ensuring a model remains "self-consistent" after an update is a major hurdle. For example, changing a head of state's name should ideally propagate to related queries about their spouse or administration, which current methods often fail to do.

Conclusion and Future Directions

While significant progress has been made, several gaps remain. There is a dire need for unified benchmarks that evaluate CL, ME, and Retrieval under identical criteria, moving beyond simple fact-checking to assess complex reasoning. Furthermore, as models become increasingly multimodal, expanding knowledge in vision-language systems represents a critical new frontier for research.

Summary updated

Abstract

Adapting large language models (LLMs) to new and diverse knowledge is essential for their lasting effectiveness in real-world applications. This survey provides an overview of state-of-the-art methods for expanding the knowledge of LLMs, focusing on integrating various knowledge types, including factual information, domain expertise, language proficiency, and user preferences. We explore techniques, such as continual learning, model editing, and retrieval-based explicit adaptation, while discussing challenges like knowledge consistency and scalability. Designed as a guide for researchers and practitioners, this survey sheds light on opportunities for advancing LLMs as adaptable and robust knowledge systems.
BibTeX
@inproceedings{wang2025bring,
  author = {Mingyang Wang and Alisa Stoll and Lukas Lange and Heike Adel and Hinrich Schütze and Jannik Strötgen},
  title = {{Bring Your Own Knowledge: A Survey of Methods for LLM Knowledge Expansion}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure