Quiet-STaR essentially trains AI systems to simulate a varied inner monologue, enabling them to anticipate various conversational trajectories and learn from ongoing interactions. This is very different from conventional AI chatbots (like ChatGPT), which lack the capacity to think about responses or foresee different conversational outcomes.
World’s Leading High-rise Marketplace
The results of applying Quiet-STaR to Mistral 7B, an open-source large language model (LLM), are impressive. Before any training, Mistral 7B scored 36.3 percent on a reasoning test. However, after implementing Quiet-STaR, its score jumped to 47.2 percent. While the AI still struggled with specific tasks, such as a school math test, its performance nearly doubled when using an inner monologue.
This post was originally published on 3rd party site mentioned in the title of this site