ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences
ScienceClaw tests whether science agents can turn verified fixes into durable program improvements.
The paper defines self-evolution as fixed-parameter program updates across task solving, verification, and workflow repair. Its evaluation spans 23 natural and social science disciplines, tracking correctness, gain, retention, transfer, and cost across sequential task streams. Updates are kept only if replay reproduces the original repair and independent scientific tasks improve. ArXiv · AI/CL/LG's note
The paper defines self-evolution as fixed-parameter program updates across task solving, verification, and workflow repair. Its evaluation spans 23 natural and social science disciplines, tracking correctness, gain, retention, transfer, and cost across sequential task streams. Updates are kept only if replay reproduces the original repair and independent scientific tasks improve. ArXiv · AI/CL/LG's note
score 5