Scientists Gave AI an “Inner Monologue” and Something Fascinating Happened

Key Points:

  • Creation of Quiet Self-Taught Reasoner (Quiet-STaR) AI model
  • Training the model to show its work before providing answers
  • Improvement in reasoning capabilities through self-teaching on diverse web text


Researchers from Stanford University and the group Notbad AI have collaborated to develop an innovative AI model called Quiet Self-Taught Reasoner (Quiet-STaR). This model distinguishes itself by pausing to “think,” displaying its reasoning process, and asking users to identify the most accurate response. Building on the previous Self-Taught Reasoner algorithm, the team sought to create an AI that emulates human inner monologues, contemplating before providing answers.


Quiet-STaR is constructed on Mistral 7B, an open-source large language model known for its seven billion parameters, exceeding the capabilities of Meta’s Llama model. The AI exhibits its reasoning behind answers to allow users to select the most precise response, resulting in a 47.2 percent accuracy rate, an improvement from its previous 36.3 percent performance. Notably, the model’s math proficiency doubled during training, showcasing potential advancements in common-sense reasoning compared to existing chatbots like OpenAI’s ChatGPT and Google’s Gemini.


While still lacking in certain areas such as mathematics, the Quiet-STaR AI presents promising progress that could bridge the gap between language models and human-like reasoning abilities. Researchers speculate that this development may pave the way for significant advancements in AI technology, potentially rivaling projects like OpenAI’s Q* model. The impact of Quiet-STaR’s unique approach to self-teaching reasoning could be transformative, with implications for enhancing AI capabilities that emulate human thought processes and decision-making.



Prompt Engineering Guides



©2024 The Horizon