
Meta's hyperagents improve at tasks and improve at improving
1 sources·1 updates
Meta's researchers, in collaboration with several universities, have introduced hyperagents—AI systems capable of not only completing tasks but also optimizing their own improvement processes. This innovative approach, known as DGM-H, allows the agents to enhance their performance across various domains, achieving significant gains in tasks such as paper review and robotics, and even transferring learned strategies to new challenges like Olympiad math evaluations.
Key Points
DGM-H achieved a performance increase from 0.0 to 0.710 in paper review tasks.
Hyperagents can transfer improvement strategies across domains, scoring 0.630 in Olympiad math evaluations.
Timeline
Get personalized news summaries delivered to your feed
Try Trace Free