Summary of Larger Language Models Don’t Care How You Think: Why Chain-of-thought Prompting Fails in Subjective Tasks, by Georgios Chochlakis et al.
Larger Language Models Don’t Care How You Think: Why Chain-of-Thought Prompting Fails in Subjective Tasks
by Georgios Chochlakis, Niyantha Maruthu Pandiyan, Kristina Lerman, Shrikanth Narayanan
First submitted to arxiv on: 10 Sep 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 In-Context Learning (ICL) in Large Language Models (LLMs) is a dominant technique for performing natural language tasks, offering competitive or state-of-the-art results at reduced computational costs. ICL can be augmented with Chain-of-Thought (CoT) prompting, which involves incorporating reasoning processes into the prompt. However, recent research suggests that ICL relies more on retrieving task priors than true learning, especially in complex domains like emotion and morality. This study investigates whether CoT also suffers from the same behavior as ICL, retrieving reasoning priors despite evidence in the prompt. The results show that CoT indeed exhibits posterior collapse for larger LLMs, highlighting the need to explore alternative approaches. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research looks at a technique called In-Context Learning (ICL) used by large language models (LLMs). ICL is good at performing natural language tasks quickly and efficiently. Some researchers added a way to make ICL even better, called Chain-of-Thought (CoT) prompting. However, recent studies found that ICL doesn’t really “learn” new things, but instead uses pre-existing knowledge. This study wanted to see if CoT has the same problem. Surprisingly, it does! The results show that CoT can also get stuck in a pattern of using prior knowledge, rather than learning from the task at hand. |
Keywords
» Artificial intelligence » Prompt » Prompting
