Can you automate the LLM fine tuning process to handle changing business laws without an expert?
Table of Contents
Tired of manual prompt updates? Discover how self-evolving AI agents learn from their own failures to automate fine-tuning and adapt to new rules on the fly.
Key Takeaways
What: Self-evolving multi-AI agents that learn and adapt autonomously.
Why: Removes the need for constant manual expert updates to prompts and business rules.
How: Agents analyze task failures to extract actionable knowledge and update their own operational logic.
Standard industry logic suggests that the path to a high-performing AI is paved with perfect data and constant human oversight. Most developers spend their time trying to prevent AI from making mistakes, assuming that every error requires a human engineer to step in and fix the underlying prompt or logic. Fujitsu is challenging this assumption with a technology that views failure not as a setback, but as the primary engine for growth.
Learning from the “Why” of Failure
When a typical AI fails to find a specific document or provides an incorrect answer, a skilled professional must manually adjust the search methods or evaluation criteria. Fujitsu’s self-evolving multi-AI agent technology removes this bottleneck. Instead of just flagging an error, these agents autonomously analyze why they failed and extract actionable knowledge to update their own operational rules.
This shift is critical for anyone looking to build multi ai agent framework systems that can survive in the real world. In corporate environments, laws change and system specifications are updated constantly. A system that relies on experts to manualy tweak prompts every time a rule changes will eventually stall. By allowing agents to verify their own operational experience, they can take over the complex tasks of adjusting prompts and updating judgment criteria that used to require a human expert.
Automating the Evolution of “Takane”
This autonomous logic is currently being applied to enhance “Takane,” a Large Language Model designed for specific business needs. Traditionally, creating a specialized model for a field like healthcare or finance requires a massive amount of manual labor to select data and adjust learning conditions. Fujitsu’s approach uses multi-agent systems to automate llm fine tuning process steps by having the agents themselves manage the data selection and evaluation cycles.
The results of this self-driven improvement are measurable. When applied across manufacturing, healthcare, and finance, the technology led to an average accuracy improvement of 28 points compared to models that hadn’t undergone this specialization. In the medical field, the agents proved they could handle messy, unstructured medical records and extract diagnostic names and treatment policies in a consistent, structured format. This demonstrates that AI can move beyond simple chat interfaces and into the role of a specialized business foundation that grows alongside its human coworkers.
Sovereign AI at the Edge
A major hurdle for many organizations is the need for privacy and speed, which often makes cloud-based AI a difficult sell. Fujitsu is working toward “Sovereign AI,” which allows these self-evolving systems to operate locally on-premises or at the edge. This means sensitive data never has to leave the company’s internal environment.
To make this practical, Fujitsu is collaborating with researchers from Carnegie Mellon University, including Graham Neubig and Tim Dettmers. They are integrating generative AI reconstruction technology to help these multi-agent systems run using significantly less memory and power. The goal is an intelligent system that can learn from on-site failures and human feedback in real-time, even in environments with limited hardware resources.
Standardizing Business Intelligence
Through the OneFujitsu initiative, this technology is being used to unify global IT and data processes. The agents aren’t just repeating searches; they are learning high-level exploration techniques from skilled humans. For example, when searching through complex design specifications for healthcare systems, the agents learned to look at peripheral documents that might seem irrelevant but actually contain vital context for the business domain.
By shifting the burden of AI maintenance from human specialists to the agents themselves, companies can address the growing shortage of AI personnel. This technology creates a foundation where the AI doesn’t just execute tasks—it learns the “tacit knowledge” of the workplace, ensuring that specialized skills are preserved and improved upon as the business environment changes.