Summary of Chain-of-thought Unfaithfulness As Disguised Accuracy, by Oliver Bentham et al.
Chain-of-Thought Unfaithfulness as Disguised Accuracy
by Oliver Bentham, Nathan Stringham, Ana Marasović
First submitted to arxiv on: 22 Feb 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 This research paper investigates the relationship between a large language model’s internal computations and the Chain-of-Thought (CoT) generations that produce its output. The authors propose a metric to measure a model’s dependence on its CoT, finding that certain models exhibit a scaling-then-inverse-scaling relationship between model size and faithfulness. They also find that smaller models are less faithful, but more accurate when normalized for bias. The study evaluates the generalizability of these findings across different families of proprietary models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This study looks at how well large language models understand what they’re doing as they generate text. It’s important to know this because we want to trust that the answers they give us are correct. Researchers found that some language models get better at understanding themselves as they get bigger, but others don’t. They also discovered that smaller models might not be very good at understanding themselves, but they’re actually pretty accurate. This research helps us understand how well language models really work. |
Keywords
* Artificial intelligence * Large language model
