The paper introduces Praxist, a framework that rethinks how autonomous R&D agents are evaluated and improved. Today, these agents write, run, and refine executable artifacts under automated evaluation, but they largely function as laboratory instruments. They are demonstrated on curated benchmarks, where improvements are hard to attribute to any particular change, and the overall costs remain poorly understood.
Praxist argues for a shift from treating each output as an isolated experimental artifact to viewing work in terms of solution lineages. This means tracking the evolution of solutions over time, making it possible to see how and why an agent's performance changes. The authors suggest this lineage-based perspective could address the traceability and cost problems that plague current approaches.
The paper does not provide specific experimental results in the abstract, but it frames a clear problem and a proposed direction. For researchers working on autonomous scientific discovery, Praxist offers a conceptual alternative to benchmark-driven evaluation, one that emphasizes provenance and cumulative progress over one-off artifacts. The full details of the framework are not available in the abstract, so the specifics of its implementation remain to be examined in the paper itself.