Summary of From Loops to Oops: Fallback Behaviors Of Language Models Under Uncertainty, by Maor Ivgi et al.
From Loops to Oops: Fallback Behaviors of Language Models Under Uncertainty
by Maor Ivgi, Ori Yoran, Jonathan Berant, Mor Geva
First submitted to arxiv on: 8 Jul 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 investigates the undesirable behaviors of Large Language Models (LLMs), specifically hallucinations and sequence repetitions. The authors propose a new perspective on these behaviors, viewing them as fallbacks that models exhibit under epistemic uncertainty. They categorize three types of fallback behaviors: sequence repetitions, degenerate text, and hallucinations. By analyzing models from the same family with varying degrees of pretraining tokens, parameter count, or instruction-following training, the authors reveal a consistent ordering of these fallback behaviors across different axes. The study shows that more advanced LLMs tend to exhibit fewer sequence repetitions and more hallucinations as uncertainty increases. Additionally, the authors demonstrate that common decoding techniques can alleviate unwanted behaviors like sequence repetitions but increase harder-to-detect hallucinations. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how big language models work sometimes. They do things they shouldn’t, like making up new information or repeating themselves. The researchers want to understand why this happens and if it’s related to the model being more advanced or not. They tested several models that are similar but have different amounts of training data or parameters. What they found is that the better the model, the more likely it is to make up information instead of repeating itself. They also showed that special techniques used to fix these problems can actually make them worse. |
Keywords
» Artificial intelligence » Pretraining
