Rohan Paul

Meta published A-MLE, a new AI agent framework they actually built and deployed to automate ML experimentation across their ads-ranking models. All because, production ML is often limited by how fast engineers can test ideas. A-MLE, Meta's automates the repetitive loop behind improving ads-ranking models: proposing ideas, running experiments, recovering failed jobs, comparing results, and carrying useful lessons to other models. A generic LLM scored 8% on Meta’s basic capability test, while the domain-equipped agent reached 68%. On 1 experimental ranking model, broader exploration cut offline regression error by 2.56% relative to baseline while training QPS stayed essentially unchanged at +0.42%.