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

📝 Paper Summary

Knowledge expansion Continual learning Model editing Retrieval-augmented generation

This survey provides a taxonomy and overview of methods for expanding LLM knowledge—categorizing techniques into continual learning, model editing, and retrieval—across factual, domain, language, and preference dimensions.

Core Problem

LLMs are typically trained once with a cutoff date, making their internal knowledge static and unable to adapt to evolving facts, specialized domains, new languages, or changing user preferences without intervention.

Why it matters:

Static models become obsolete as real-world information changes (factual decay), limiting their utility in time-sensitive applications
General-purpose models often fail in specialized fields like medicine or law without targeted domain adaptation
Re-training fully is computationally prohibitive, creating a need for efficient adaptation strategies that mitigate catastrophic forgetting

Concrete Example: An LLM trained in 2021 will not know about the 2023 Nobel Prize winners. Without knowledge expansion, it hallucinates or refuses to answer. A retrieval-based method would fetch the news, while model editing would directly modify the weights associated with 'Nobel Prize 2023'.

Key Novelty

Task-Oriented Knowledge Expansion Taxonomy

Classifies expansion methods not just by technique (continual learning vs. editing vs. retrieval) but by the *type* of knowledge being integrated (factual, domain, language, preference)
Contrasts 'implicit' knowledge expansion (modifying internal parameters via continual learning or editing) with 'explicit' expansion (retrieval-based access during inference) to guide selection based on use-case needs

Evaluation Highlights

Review of Continual Pretraining (CPT) effectiveness for language expansion, citing Glot500 extending support to 500 languages
High-level comparison of methods showing retrieval is best for 'Plug-and-Play' flexibility, while Continual Learning excels at 'Generalization' (Table 1 summary)
Survey of programming language expansion showing domain-specific models like CodeLLaMA and StarCoder 2 consistently outperform general-purpose LLMs on code benchmarks

Breakthrough Assessment

7/10

A comprehensive survey that structures a fragmented field. While it doesn't propose a new model, its clear taxonomy (Facts/Domain/Language/Preference × CL/Editing/Retrieval) is a valuable contribution for researchers navigating adaptation choices.

⚙️ Technical Details

Problem Definition

Setting: Adapting a pre-trained Large Language Model to new knowledge sets without full retraining

Inputs: A pre-trained LLM and a new knowledge source (text corpus, factual triples, preference pairs, or external document index)

Outputs: An adapted LLM (updated weights) or an augmented inference system capable of utilizing the new knowledge

Pipeline Flow

Select Knowledge Type (Factual, Domain, Language, Preference)
Select Method Strategy (Implicit vs. Explicit)
Apply Method (Continual Learning, Model Editing, or Retrieval)

System Modules

Continual Learning (CL) (Implicit Expansion)

Update parameters incrementally using new corpora

Model or implementation: Various (e.g., DEMix-DAPT, LLaMAX, CPPO)

Model Editing (Implicit Expansion)

Precise modification of specific knowledge units

Model or implementation: Editor networks (e.g., MEND, ROME - implied context)

Retrieval-Based

Fetch external context at inference time

Model or implementation: Retriever + Generator

Novel Architectural Elements

Taxonomy matrix integrating 4 knowledge types (Factual, Domain, Language, Preference) with 3 adaptation strategies (CL, Editing, Retrieval)

Modeling

Base Model: Survey covers multiple models (e.g., LLaMA, XLM-R, mT5)

Training Method: Survey of various methods: Continual Pretraining (CPT), Continual Preference Alignment (CPA), Adapter-based tuning

Objective Functions:

Purpose: Prevent catastrophic forgetting during continual learning.

Formally: Regularization loss (e.g., penalizing changes to important parameters) or Replay loss (mixing old data)
Purpose: Align with evolving user preferences.

Formally: CPPO (Continual Proximal Policy Optimization) using sample-wise weighting strategies

Adaptation: Includes parameter-efficient methods (LoRA, adapters) and full fine-tuning strategies

Trainable Parameters: Varies by method (from sparse updates in Model Editing to full updates in CPT)

Training Data:

Domain-specific corpora (e.g., medical, legal)
Multilingual datasets (e.g., Glot500 corpus)
Preference pairs for alignment

Compute: Not reported in the paper

Comparison to Prior Work

vs. Static Training: Knowledge expansion methods allow adaptation after the cutoff date without full re-training
Implicit (CL, ME) vs. Explicit (RAG): Implicit methods modify parameters for internalized knowledge; Explicit methods rely on external buffers, reducing parameter dependency
Structured Taxonomy [Novelty]: This survey systematically maps adaptation methods to specific knowledge types (Fact, Domain, Language, Preference), whereas prior surveys often focus on just one (e.g., temporal factual updates or editing)

Limitations

Continual Learning suffers from computational inefficiency and lack of precise control compared to editing
Model Editing is primarily suited for factual updates and struggles with broad domain or language adaptation
Retrieval-based methods depend heavily on the quality of the external index and retrieval accuracy
Survey focuses on textual LLMs, less emphasis on multimodal knowledge expansion

Reproducibility

Survey paper; does not provide a single code repository. References external repositories for specific methods discussed (e.g., Glot500, DEMix-DAPT).

📊 Experiments & Results

Evaluation Setup

Survey of existing literature results across multiple domains

Benchmarks:

Glot500 (Multilingual Language Modeling)
CodeTask-CL (Code Generation/Summarization)
Aya (Multilingual Instruction Tuning)

Metrics:

Accuracy (factual)
Perplexity (language modeling)
Forgetting Rate (continual learning)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
The survey aggregates qualitative and quantitative findings from cited papers rather than running new experiments. Key cited results include:

Main Takeaways

No single method fits all knowledge types: Continual Learning is best for broad Language/Domain adaptation, while Model Editing is best for specific Factual corrections.
Retrieval offers the best 'freshness' for rapidly changing facts but doesn't improve the model's intrinsic reasoning or language capabilities.
Continual Preference Alignment (CPA) is an emerging critical area, shifting from static RLHF to dynamic alignment with evolving societal values.
Parameter isolation techniques (e.g., MoE, adapters) are increasingly standard to mitigate catastrophic forgetting during domain expansion.

📚 Prerequisite Knowledge

Prerequisites

Understanding of Large Language Model pre-training vs. fine-tuning
Familiarity with catastrophic forgetting in neural networks
Basic knowledge of RAG (Retrieval-Augmented Generation) and parameter-efficient fine-tuning (PEFT)

Key Terms

Continual Learning: A paradigm where models incrementally learn from new data streams without forgetting previously learned information

Model Editing: Techniques for precisely modifying specific facts or behaviors in a model's weights without retraining the entire network (e.g., locating and changing a specific neuron)

Retrieval-Augmented Generation: RAG—Providing external knowledge to a model during inference by searching a database, rather than baking it into the model weights

Catastrophic Forgetting: The tendency of neural networks to lose previously learned information when trained on new data

Instruction Tuning: Fine-tuning a model on datasets of instructions and responses to improve its ability to follow user commands

RLHF: Reinforcement Learning from Human Feedback—aligning models to human values using reward models trained on preference data

Parameter-Isolation: A continual learning strategy that allocates different parameter subsets to different tasks or domains, keeping the majority frozen to prevent forgetting

Self-Information Updating: A method where the model generates its own training data based on new information to update its internal knowledge