Hyperagents — New Continuously Self-Improving AI

Researchers at Meta AI, working alongside teams from several universities and research labs, have introduced Hyperagents — a new approach to building AI systems that not only improve task performance but also refine their own learning mechanisms.

At the core of the concept is the combination of two roles within a single agent: a task executor and an internal evaluator that analyzes outcomes and adjusts the process used to achieve them. Unlike traditional models that rely on fixed self-learning rules, Hyperagents can adapt the learning process itself. The researchers describe this as metacognitive self-modification.

The approach builds on the Darwin Gödel Machine (DGM) concept but removes its reliance on predefined instructions for self-improvement. In the updated version, DGM-Hyperagents (DGM-H), the optimization process becomes more flexible and is no longer tied to a specific domain, making it applicable across a wide range of computational tasks.

In experiments, the system demonstrated steady improvements across multiple domains, including programming, scientific paper analysis, mathematical problem solving, and robotics algorithm design. At the same time, Hyperagents continuously refine their own learning strategies by accumulating experience, tracking the effectiveness of solutions, and adapting their methods.

The self-improvement capabilities developed in one domain transfer to others, increasing the system’s overall versatility. According to the researchers, this approach points toward AI systems that can continuously improve how they learn, potentially accelerating technological progress.