A new arXiv preprint applies Andrej Karpathy's AutoResearch paradigm to a costly production problem: optimizing embedding systems for recommendation pipelines. The authors argue that systematic exploration in this area consumes a disproportionate amount of engineering effort at scale.
The paper, titled "AutoResearch at Production Scale," focuses on the failure modes that emerge when this paradigm is scaled. To address them, the authors introduce a multi-agent framework designed to reduce manual burden and make the exploration process more systematic.
Because this is a preprint, the findings have not yet been peer-reviewed. The abstract is brief, so the specific failure modes and framework details are left to the full paper.