DMC-Optim: A New Benchmark for Training AI to Write Faster Code

burkov · x · 2026-08-03

Current code generation models often focus solely on correctness at the expense of execution efficiency. Simply adding runtime to the reward during reinforcement learning frequently fails due to noisy timing measurements and unstable updates from sparse feedback.

To address this, researchers from Inria, Meta, and Paris-Dauphine introduced DMC-Optim, a benchmark built on competitive programming problems. It uses stronger correctness checks and larger inputs to reveal meaningful speed differences among correct solutions. By combining isolated runtime measurement, speed rewards granted only to correct code, a cheap simulator for hyperparameter selection, and strategies to stabilize model updates, they significantly improved the model's ability to generate efficient code.

Original post →

More from Research

Research channel →