CORAL: An LLM-Native Harness for Production Recommender Systems
CORAL puts an LLM agent in the loop to keep tuning live recommender systems under fixed guardrails.
The paper frames recommender maintenance as a partially observed, non-stationary constrained optimization problem. In each cycle, the agent reads operating signals, uses memory of prior changes, and calls tools including a numerical optimizer to reconfigure retrieval, ranking, or serving choices. A/B tests on two large social platforms found engagement gains at no extra serving cost on one platform, and lower serving cost without engagement loss on the other. The authors say performance improved over repeated loops, pointing to automation of work normally handled by human algorithm engineers. ArXiv · AI/CL/LG's note
The paper frames recommender maintenance as a partially observed, non-stationary constrained optimization problem. In each cycle, the agent reads operating signals, uses memory of prior changes, and calls tools including a numerical optimizer to reconfigure retrieval, ranking, or serving choices. A/B tests on two large social platforms found engagement gains at no extra serving cost on one platform, and lower serving cost without engagement loss on the other. The authors say performance improved over repeated loops, pointing to automation of work normally handled by human algorithm engineers. ArXiv · AI/CL/LG's note
score 5