Summary of Wise: Rethinking the Knowledge Memory For Lifelong Model Editing Of Large Language Models, by Peng Wang et al.
WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models
by Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, Huajun Chen
First submitted to arxiv on: 23 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Information Retrieval (cs.IR); 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 Medium Difficulty summary: Large language models (LLMs) require knowledge updates to address the increasing gap between world facts and their responses, enabling lifelong model editing. The study investigates where updated knowledge resides in memories and proposes a novel approach called WISE to bridge this gap. The authors find that editing either long-term memory (direct model parameters) or working memory (non-parametric neural network activations/representations by retrieval) results in an “impossible triangle” – reliability, generalization, and locality cannot be achieved together. They design a dual parametric memory scheme with a main memory for pretrained knowledge and a side memory for edited knowledge, using a router to decide which memory to access. A knowledge-sharding mechanism is also proposed for continual editing. The authors demonstrate the effectiveness of WISE through extensive experiments on question answering, hallucination, and out-of-distribution settings across various LLM architectures, including GPT, LLaMA, and Mistral. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Low Difficulty summary: This paper is about how to improve large language models so they can learn new things and correct mistakes. The authors want to know where this new knowledge goes in the model’s memory. They found that there are some limitations when trying to update the model’s memories, making it hard to achieve three important goals – being reliable, generalizable, and local. To solve this problem, the authors created a new way of updating the model’s memories called WISE. This approach uses two separate areas in the memory: one for the original knowledge and one for the updated knowledge. The authors tested WISE and showed that it works better than other methods. |
Keywords
» Artificial intelligence » Generalization » Gpt » Hallucination » Llama » Neural network » Question answering
