Summary of Quiet-star: Language Models Can Teach Themselves to Think Before Speaking, by Eric Zelikman et al.
Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking
by Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, Noah D. Goodman
First submitted to arxiv on: 14 Mar 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 paper presents Quiet-STaR, a novel approach for language models (LMs) to learn reasoning skills by generating rationales at each token to explain future text. The authors build upon the Self-Taught Reasoner (STaR) framework, which has shown promise in learning from few-shot examples and inferring unstated rationales in question-answering and learning settings. Quiet-STaR addresses key challenges, including computational cost, lack of initial knowledge on generating internal thoughts, and predicting beyond individual next tokens. To resolve these, the authors propose a tokenwise parallel sampling algorithm using learnable tokens indicating a thought’s start and end, and an extended teacher-forcing technique. The approach demonstrates zero-shot improvements on GSM8K (5.9% → 10.9%) and CommonsenseQA (36.3% → 47.2%), as well as perplexity improvement of difficult tokens in natural text, without requiring fine-tuning on these tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps computers become better at understanding what we mean when we write or talk. Right now, computers can answer simple questions by looking at the words you type, but they struggle to understand complex thoughts and ideas. The authors want to change this by teaching computers to think more like humans. They’ve developed a new way for computers to learn from a small number of examples and generate explanations for why they made certain decisions. This approach shows promise in improving computer’s ability to answer difficult questions and understand natural language. |
Keywords
* Artificial intelligence * Few shot * Fine tuning * Perplexity * Question answering * Token * Zero shot
