- March 29, 2024
- Research
A partnership between artificial intelligence experts from Stanford University and Notbad AI Inc. led to the creation of an innovative algorithm that allows modern chatbots to consider different answers to a question before presenting a final answer. The group shared their findings in a detailed report on the arXiv preprint server, demonstrating the effectiveness of their method when integrated with an already existing chatbot.
Experts emphasize that today most chatbots generate answers directly based on training data and do not consider different possible answers before providing the most likely one. This type of instant response will be similar to the way a person hastily blurts out the first thing that comes to mind without thinking it through.
In a recent project, the team equipped chatbots with the ability to think before they respond, thus significantly increasing their accuracy and making responses more human-like. Their solution, called Quiet-STaR, initially prompts the chatbot to offer multiple responses to a query. It then compares them to the original question to determine the most appropriate answer, which is then relayed to the user. Additionally, the algorithm is designed to improve the scoring process over time by learning past interactions.
Quiet-STaR’s performance was evaluated by integrating it with the open-source Mistral 7B chatbot for a standard reasoning test, in which it scored 47.2%, in stark contrast to Mistral 7B’s non-algorithm score of 36.3%. He also showed significant improvements in math scores. The researchers suggest that their algorithm could improve the performance of the various chatbots currently available, provided their developers implement it. This innovation aims to improve the overall accuracy of chatbots.
