CORAL Puts an LLM Agent in a Live Closed Loop on Production Recommenders and Wins Two A/B Tests
Most LLM work on recommenders touches ranking, user modeling or offline development; CORAL instead runs an agent in a continual loop against a live system, observing operating signals each cycle, reasoning over a memory of past decisions and their measured outcomes, and calling tools including a numerical optimizer that holds changes inside a fixed operating budget. It is framed as a partially observed, non-stationary, constrained optimization problem where the policy improves in context from its own prior actions, with no parameter updates. Evaluated by A/B experiment on two large-scale social platforms, the same harness improved engagement at no additional serving cost on one and reduced serving cost without degrading engagement on the other.
↳ Follow the thread