QuietSTaR: AI’s Inner Monologue Enhances Reasoning and Learning
Scientists have developed QuietSTaR, a method that trains AI systems to think before responding, akin to human inner monologue. This technique significantly improves common sense reasoning and math performance. Unlike traditional AI chatbots, QuietSTaR-equipped AI generates multiple inner rationales before responding, learning from mistakes to anticipate and adapt to future conversations.
Applied to Mistral 7B, an LLM, QuietSTaR boosted reasoning scores by 10.9% and doubled math performance. The approach marks a departure from domain-specific methods, promising broader applicability in enhancing AI’s reasoning capabilities and bridging the gap between neural network-based AI and human-like reasoning.