Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking logo

Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking

Free

Language Models Can Teach Themselves to Think Before Speaking

FreeFree tier
Inputs: textOutputs: text
Type
Open Source

About Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking

Quiet-STaR is a generalization of the Self-Taught Reasoner (STaR) that enables language models to learn to generate internal rationales—'thoughts'—at each token position to explain and predict future text, improving their overall predictions. It addresses key challenges such as the high computational cost of generating continuations, the model's initial inability to generate or use internal thoughts, and the need to predict beyond individual next tokens. The method introduces a tokenwise parallel sampling algorithm, learnable start and end tokens that demarcate thoughts, and an extended teacher-forcing technique. After continued pretraining on a corpus of internet text, Quiet-STaR yields zero-shot improvements on GSM8K (from 5.9% to 10.9%) and CommonsenseQA (from 36.3% to 47.2%) without any fine-tuning on these tasks, along with reduced perplexity on difficult tokens in natural text. This approach marks a step toward more general and scalable reasoning in language models.

Key Features

Tokenwise parallel sampling algorithm for generating rationales at each token
Learnable start and end tokens to indicate the beginning and end of internal thoughts
Extended teacher-forcing technique to handle prediction beyond next tokens
Zero-shot improvements on reasoning benchmarks without task-specific fine-tuning
Generalization of STaR from question-answering to arbitrary natural text

Pros & Cons

Pros
  • Improves prediction of difficult-to-predict tokens in natural text
  • Achieves zero-shot improvements on reasoning tasks without fine-tuning
  • Provides a general and scalable framework for learning to reason from arbitrary text
  • Addresses key challenges in generating and using internal thoughts during autoregressive generation
Cons
  • High computational cost due to generating continuations and thoughts at every token
  • Requires continued pretraining on a large corpus of internet text for effective learning
  • May not be directly applicable to all language model architectures or domains without adaptation

Best For

Improving language model perplexity on difficult-to-predict tokens in arbitrary textZero-shot reasoning on math word problems (GSM8K)Zero-shot commonsense reasoning (CommonsenseQA)Enabling language models to infer unstated rationales in conversations, proofs, and other implicit reasoning contexts

FAQ

What is Quiet-STaR?
Quiet-STaR is a generalization of the Self-Taught Reasoner (STaR) that allows language models to learn to generate internal rationales (thoughts) at each token position to explain and improve predictions of future text.
How does Quiet-STaR work?
It uses a tokenwise parallel sampling algorithm to generate rationales, learnable start and end tokens to demarcate thoughts, and an extended teacher-forcing technique to train the model to use these thoughts for better prediction.
What benchmarks does Quiet-STaR improve?
After continued pretraining on internet text, Quiet-STaR shows zero-shot improvements on GSM8K (math reasoning) from 5.9% to 10.9% and on CommonsenseQA from 36.3% to 47.2% without any fine-tuning.