Fujitsu AI Learns from Mistakes: 28% Accuracy Gain

Fujitsu has developed a new self-evolving multi-AI agent system. These AI teams can continuously improve their own performance by learning from operational results, human feedback, and other changes. Here's the thing: the AI agents can identify why they succeed or fail. They then extract insights and update their own prompts, search strategies, and evaluation criteria. This means tasks that used to need human experts are now handled by the AI itself. Fujitsu tested this technology in two areas. When applied to their large language model, Takane, the system managed data selection and training. It delivered an average accuracy gain of 28 points. It was also used for design specification searches, where agents learned from past failures to refine document retrieval. Fujitsu plans to integrate this technology into its Kozuchi AI platform. They are also working with Carnegie Mellon University to develop versions for confidential infrastructure settings. This could mean more efficient and adaptable AI systems across many industries.