LLM Surgery: Efficient Knowledge Unlearning and Editing in Large Language Models (2409.13054v1)

Abstract: LLMs have revolutionized various domains, yet their utility comes with significant challenges related to outdated or problematic knowledge embedded during pretraining. This paper addresses the challenge of modifying LLMs to unlearn problematic and outdated information while efficiently integrating new knowledge without retraining from scratch. Here, we propose LLM Surgery, a framework to efficiently modify LLM behaviour by optimizing a three component objective function that: (1) Performs reverse gradient on unlearning dataset (problematic and outdated information), (2) Performs gradient descent on the update dataset (new and updated information), and (3) Minimizes the KL divergence on the retain dataset (small subset of unchanged text), ensuring alignment between pretrained and modified model outputs. Due to the lack of publicly available datasets specifically tailored for our novel task, we compiled a new dataset and an evaluation benchmark. Using Llama2-7B, we demonstrate that LLM Surgery can achieve significant forgetting on the unlearn set, a 20\% increase in accuracy on the update set, and maintain performance on the retain set.

• The paper introduces LLM Surgery, an efficient framework using a three-pronged objective function with reverse gradient for unlearning, gradient descent for updating, and KL divergence for retaining knowledge in LLMs.
• Experiments show LLM Surgery effectively reduces obsolete knowledge and improves accuracy on new data while preserving existing knowledge on Llama2-7B, achieving a 35x reduction in GPU hours compared to naive retraining.
• This efficient method helps LLMs remain current, accurate, and compliant with regulations like the Right to be Forgotten, addressing vital real-world challenges related to dynamic information.

LLM Surgery: Efficient Knowledge Unlearning and Editing in LLMs

The paper "LLM Surgery: Efficient Knowledge Unlearning and Editing in LLMs" presents a comprehensive framework designed to update and optimize the knowledge embedded within LLMs efficiently. The authors focus on modifying LLMs to unlearn obsolete or erroneous information and assimilate new knowledge without necessitating a full retrain of the model. This paper introduces and evaluates an innovative approach termed LLM Surgery, which addresses the inherent challenges associated with maintaining the relevance and accuracy of LLMs over time.

Key Contributions and Methodology

LLM Surgery employs a three-pronged objective function to optimize LLM behavior. This includes:

1. Unlearning Mechanism: Implementing reverse gradient on the unlearning dataset, which comprises problematic and outdated information. This component effectively reduces the model's likelihood of reproducing incorrect or sensitive content.
2. Knowledge Update: Utilizing gradient descent on the update dataset, which involves integrating new and relevant data into the model. This ensures that the LLM can provide updated and accurate information.
3. Retention of Essential Knowledge: Minimizing the KL divergence on the retain dataset. This ensures the modified LLM maintains alignment with the original model's outputs where necessary, preserving essential knowledge.

The authors compiled a novel dataset due to the absence of publicly available datasets for such tasks and designed a specific evaluation benchmark to measure the efficacy of these components.

Experimental Setup and Results

The experiments utilize the Llama2-7B model as the base for testing the framework's effectiveness. The paper presents empirical evidence showcasing that LLM Surgery can significantly improve the model's ability to forget obsolete information and incorporate new knowledge while maintaining stable performance on preexisting tasks. Key results include:

• A reduction in accuracy on the unlearn set, demonstrating effective knowledge forgetting.
• A 20% improvement in accuracy on the update set, indicating successful integration of new data.
• Stable performance on tasks represented by the retain set, affirming the preservation of necessary knowledge without unwanted degradation.

Crucially, LLM Surgery achieves these outcomes while significantly reducing computational overhead compared to naive retraining approaches, demonstrating a 35x reduction in GPU hours.

Implications and Future Directions

This paper addresses vital challenges associated with the dynamic and evolving nature of information in LLMs, thus preventing issues related to outdated or legally problematic data. The approach is scalable and applicable to unstructured data, which is pivotal for real-world implementations and regulatory compliance, including adherence to laws like the Right to be Forgotten.

Potential future developments may include refining the balance between unlearning and retaining data or extending the framework to support other model architectures. The authors' proposal to publicly release their code and datasets may catalyze further research, potentially leading to improved methodologies and broader community engagement in updating LLM behavior effectively.

In essence, the LLM Surgery framework is a significant step forward in allowing LLMs to remain accurate, current, and legal in the ever-changing landscape of information.