Summary of The Butterfly Effect Of Model Editing: Few Edits Can Trigger Large Language Models Collapse, by Wanli Yang et al.
The Butterfly Effect of Model Editing: Few Edits Can Trigger Large Language Models Collapse
by Wanli Yang, Fei Sun, Xinyu Ma, Xun Liu, Dawei Yin, Xueqi Cheng
First submitted to arxiv on: 15 Feb 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: None
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 This paper investigates the impact of model editing on Large Language Models (LLMs), a crucial aspect often overlooked despite its promise in revising knowledge. The authors reveal a critical phenomenon: single edits can trigger model collapse, resulting in significant performance degradation across various benchmark tasks. To mitigate this, they propose using perplexity as a surrogate metric, validated through extensive experiments demonstrating strong correlations with downstream task performances. The paper further explores sequential editing, a practical setting for real-world scenarios, examining various editing methods and LLMs. The results show that nearly all examined editing methods lead to model collapse after only a few edits. To facilitate research, the authors developed a new dataset, HardEdit, based on these hard cases, aiming to establish a foundation for pioneering studies in reliable model editing and mechanisms underlying editing-induced model collapse. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at what happens when you make changes to Large Language Models (LLMs). Researchers found that even small edits can cause the model to break down and perform poorly. To avoid this, they suggest using a new metric called perplexity, which helps predict how well the model will do in different tasks. They also tested making multiple changes at once and found that most editing methods cause the model to collapse after just a few edits. The authors created a new dataset to help other researchers study these issues and find ways to make reliable changes to LLMs. |
Keywords
» Artificial intelligence » Perplexity
