Experiment trains a 4B open model to generate Postgres query plans with 44.7% lower latency on join-heavy workloads

A researcher post-trained a 4B open-weights model using supervised fine-tuning and reinforcement learning to produce Postgres query plans that outperform the database’s default optimizer. On 113 join-heavy queries, the model achieved a 44.7% latency reduction, despite initially failing to produce valid plans for most of them.

The work suggests small language models can learn narrowly scoped database optimization tasks where outputs are easy to verify through execution time. If replicated, this points to a broader trend of AI-assisted systems tuning complementing traditional hand-built optimizers in complex infrastructure.