Summary of Llm Surgery: Efficient Knowledge Unlearning and Editing in Large Language Models, by Akshaj Kumar Veldanda et al.
LLM Surgery: Efficient Knowledge Unlearning and Editing in Large Language Models
by Akshaj Kumar Veldanda, Shi-Xiong Zhang, Anirban Das, Supriyo Chakraborty, Stephen Rawls, Sambit Sahu, Milind Naphade
First submitted to arxiv on: 19 Sep 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper proposes a framework called LLM Surgery to modify large language models (LLMs) by optimizing a three-component objective function. This framework aims to efficiently remove outdated or problematic knowledge embedded during pretraining while integrating new information without retraining from scratch. The approach involves reverse gradient on an unlearning dataset, gradient descent on an update dataset, and minimizing the KL divergence on a retain dataset. To evaluate this novel task, the authors compiled a new dataset and benchmark. Using Llama2-7B, they demonstrate that LLM Surgery can achieve significant forgetting on the unlearn set, improve accuracy by 20% on the update set, and maintain performance on the retain set. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models have changed many fields, but they also contain old or bad information from when they were first trained. This paper wants to fix that problem. They created a new way to change LLMs called LLM Surgery. It’s like surgery for AI models! The new method uses three parts: one to remove the bad information, one to add new information, and one to make sure the model doesn’t forget important things. To test this idea, they made a special dataset and showed that it can work. |
Keywords
» Artificial intelligence » Gradient descent » Objective function » Pretraining
