Quiet-STaR method enhances AI reasoning abilities significantly
Livescience.com — March 20, 2024, 01:00 PM UTC
Training AI models with an "inner monologue" using the Quiet-STaR method improved reasoning abilities. Mistral 7B's performance on a reasoning test increased from 36.3% to 47.2%. The AI still struggled with a math test but improved from 5.9% to 10.9%. Quiet-STaR allows AI to generate rationales before responding, enhancing its capacity to anticipate and learn from conversations. Researchers aim to bridge the gap between AI and human-like reasoning.