Another brilliant paper from Meta.

This one is worth reading if you work on production-grade ranking or recommendation systems.

Meta deployed an autonomous agent that runs the ML iteration cycle across a portfolio of production ads ranking models.

Modern ads ranking is limited by how many research, implement, train, debug, evaluate and launch cycles engineers can run, rather than by model capacity or training compute.

Each cycle takes days to weeks of senior engineer attention per model, so techniques proven on one model spread slowly to the rest.

A-MLE splits the cycle into five stages covering hypothesis generation, exploration strategy, experiment execution, result analysis, and a shared knowledge substrate.

One agent orchestrates them, calling domain-specific skills and workflows against a sandboxed execution layer, with human checkpoints at every stage boundary.

They evaluate deployment along three tiers of tool availability, autonomous workflow execution, and open-ended exploration.

They also ran a controlled cross-LLM study with the agent loop held fixed. The Claude Sonnet, Gemini and GPT families differ in execution reliability and exploration aggressiveness.

Paper: https://academy.dair.ai/papers/agentic-ml-exploration-a-mle-for-ads-ranking-2609.08248