Training a 4B model to produce 81% faster query plans than Postgres

An independent researcher trained a 4B parameter Qwen model with agentic reinforcement learning against a live Postgres measurement rig, having it propose query hints and then measuring the actual execution time of each proposal rather than relying on a static cost model. Across 113 join heavy queries from the Join Order Benchmark, the trained model's hints cut latency by up to 44.7% compared to Postgres's own query planner. Query planning is a narrow, well studied problem where small models trained against a real feedback loop can sometimes beat a mature general purpose system, because the model only needs to learn patterns specific to the workload it sees during training rather than handle every possible query shape. The result does not mean the model outperforms Postgres broadly. It was trained and measured on one benchmark's query distribution, and hint based approaches carry known risks if the underlying data or query patterns shift after training. Still, the technique is a reusable pattern for anyone willing to build a similar measurement rig: use a small model as a learned optimizer for narrow, measurable subproblems where the cost of running many candidate proposals is cheap. That pattern extends well beyond databases.