Evaluating Model Editing Techniques Across Architectures

CSA - E0 334 - 2024 | Term Project

This project investigates model editing techniques for Large Language Models (LLMs), focusing on how these models can be modified in specific domains without compromising overall performance.

Objectives

1. Evaluate Model Editing using different metrics across different architectures
2. Determine the model-independence of editing techniques

Key Metrics for Evaluation

• Generalization: How well the model answers queries related to edited facts
• Locality: Ensuring unedited queries remain unaffected
• Accuracy: Correctness of answers for edited facts
• Portability: Ability to handle multi-hop questions

Background

Large Language Models (LLMs) and Knowledge Representation

Large Language Models (LLMs) store knowledge in their parametric memory through complex neural network weights. However, updating this knowledge is challenging due to:

• High computational cost of full model retraining
• Risk of catastrophic forgetting
• Maintaining model performance across different domains

Model Editing: Conceptual Framework

Model editing aims to modify specific factual knowledge in LLMs without:

• Complete model retraining
• Significant performance degradation
• Disrupting existing knowledge representations

Fundamental Approaches to Model Editing

1. Preservation-Based Methods

   • Memory-Based Approaches

     • Store edit examples separately
     • Guide model output for specific inputs
     • Minimize direct parameter modification
     • Examples: SERAC, MemPrompt, IKE

   • Additional Parameter Techniques

     • Introduce small, trainable parameter sets
     • Focused control over specific edits
     • Minimize impact on core model
     • Examples: T-Patcher, CaliNET, GRACE

2. Direct Parameter Modification Methods

   • Locate-Then-Edit Approaches

     • Identify specific parameters related to target knowledge
     • Directly modify identified parameters
     • Precise but potentially invasive
     • Examples: ROME, MEMIT, PMET

   • Meta-Learning Techniques

     • Use a hypernetwork as a "teacher"
     • Learn optimal parameter update strategies
     • Minimize negative side effects
     • Examples: MEND, KE

Methodology

Datasets Used

• ZsRE
• COUNTERFACT
• MQuAkE (for multi-hop prompts)

Model Editing Techniques Explored

Preservation Methods

• Memory-based Models (e.g., SERAC, MemPrompt)
• Additional Parameters Techniques (e.g., T-Patcher, CaliNET)

Modification Methods

• Locate-Then-Edit Approaches (e.g., ROME, MEMIT)
• Meta-learning Techniques (e.g., MEND)

Key Findings

Results for GPT-2 XL

Technique	Generalization	Locality	Accuracy
ROME	96.5	75.41	100
MEND	58.3	44.88	94.2

Results for Llama-2 (7B Parameters)

Technique	Generalization	Locality	Accuracy
ROME	92.45	87.04	99.63
MEND	94.24	90.27	97.04

Key Observations

• Models with higher parameters perform better with editing techniques
• Multi-hop prompting remains a challenging area for most LLMs
• ROME showed superior generalization compared to MEND

Practical Implications

• Model editing offers a computationally efficient alternative to full retraining
• Potential to improve LLM performance in specific domains