It is commonly believed that the ability to think before saying something is inherent only in people, and even then not at all. However, researchers have proven that modern AI models can learn this invaluable skill as well – and called it an “internal monologue”. The new Quiet-STaR algorithm instructs the AI system to generate many internal arguments in parallel before responding. By answering prompts, the AI generates different options and deduces the best answer, and learns by discarding those formulations that turned out to be incorrect. This method allows AI models to predict further dialogues and learn from current ones.

Researchers from Stanford University and Notbad AI tested the Quiet-STaR algorithm on Mistral 7B, a large open-source language model. After training on the Quiet-STaR, the Mistral 7B version scored 47.2% on the argumentation test compared to 36.3% before the training, and improved its math test scores from 5.9% to 10.9%. Yes, I still didn’t pass the Mistral 7B test. However, the result has almost doubled, and this is just the beginning.

AI models like ChatGPT and Gemini don’t correlate the data they use with context and logic, and simply generate words without understanding the answers. The Quiet-STaR algorithm is the first successful attempt to teach AI to respond to queries by “reasoning”. Now scientists want to investigate how such algorithms can reduce the gap between AI systems based on neural networks and human reasoning ability.

Літературний редактор.

Commentary