Auto-RecSys automates research on Meta's recommendation models, a post says

A post describing a Meta paper says Auto-RecSys conducts autonomous research on recommendation models where one training run can take days. According to the post, it runs experiments in parallel across servers and keeps shared memory so work survives failures and new sessions. Natural-language skill files guide reasoning, while deterministic scripts handle operations. The post describes two improvement loops: model-specific playbooks record failed attempts and retain working pipelines, while experimental results guide the next round of ideas. It says major fixes per iteration fell from 4.0 to 1.3 as the playbook matured.