Summary of Muscle: a Model Update Strategy For Compatible Llm Evolution, by Jessica Echterhoff et al.
MUSCLE: A Model Update Strategy for Compatible LLM Evolution
by Jessica Echterhoff, Fartash Faghri, Raviteja Vemulapalli, Ting-Yao Hu, Chun-Liang Li, Oncel Tuzel, Hadi Pouransari
First submitted to arxiv on: 12 Jul 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 A novel problem is identified in the update process of Large Language Models (LLMs), where instance-level degradation (instance regression) occurs, causing users’ mental models of a language model’s capabilities to degrade. This phenomenon, known as model update regression, can lead to dissatisfaction and decreased performance on downstream tasks. The study finds that fine-tuned user-facing adapters experience negative flips when the base model is updated, even when the task training procedures remain identical. To address this issue, the authors propose a training strategy involving a compatibility adapter to minimize instance regression in model updates. This approach is shown to reduce negative flips by up to 40%, for example, when updating Llama 1 to Llama 2. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Language Models (LLMs) are updated regularly to improve their performance. However, this update process often causes the language model’s capabilities to degrade, making it harder for users to understand what the model can do. This is a problem because users have to adjust their expectations with every update. The study found that when LLMs are updated, fine-tuned adapters that were working correctly before become incorrect again. To fix this, the authors suggest training an adapter that helps the language model stay compatible with its previous versions. This approach can reduce the number of incorrect predictions by up to 40%. |
Keywords
» Artificial intelligence » Language model » Llama » Regression
