ADAPT for Science: A Multi-Agent System for Self-Evolving Scientific Optimization

Key points

ADAPT split scientific optimization into two decisions: it retrieved 64 candidate initializations from a library of 15,409 historical designs to decide where to search, and evolved the optimization procedure from execution feedback to decide how to search.
Through evolution, the start-selection stage discovered a progressive 64 →8 →2 →1 budget allocation and reduced the selection policy's mean loss to 3.0054.
The optimizer-evolution stage kept the Adam structure but scaled each update by s(r) = 0.05 + 0.95r2 according to the remaining budget and skipped non-finite evaluations.
On eight unseen topologies ADAPT achieved a mean loss of 0.204±0.101, reducing loss by 54.9% versus the strongest classical optimizer and by 62.2% versus the strongest general-purpose agentic evolution baseline.
In a representative case, standard Adam reached a loss of 0.666 and ADAPT reached 0.155, a 76.7% reduction.
The authors state the limitation that the study focuses on gravitational-wave detector design and depends on the quality and coverage of the experience library.

External research summaries. These are not HDATF publications or measured product results.