This paper presents AlphaDev, a deep reinforcement learning agent that discovers novel sorting algorithms surpassing human-designed benchmarks. Formulated as a single-player game, AlphaDev learns to generate efficient assembly code, resulting in algorithms integrated into the LLVM standard C++ sort library. This demonstrates AI's potential to optimize fundamental algorithms beyond the capabilities of current human expertise.

Daniel J. Mankowitz, Andrea Michi, Anton Zhernnov, Marco Gelmi, Marco Selvi, Cosmin Paduraru, Edouard Laurent, Shariq Iqbal, Jean-Baptiste Lespiau, Alex Ahern, Thomas Köppe, Kevin Millikin, Stephen Gaffney, Sophie Elster, Jackson Broshear, Chris Gamble, Kieran Milan, Robert Tung, Minjae Hwang, Taylan Cemgil, Mohammadinam Barekatain, Yuji Li, Amol Mandhane, Thomas Hubert, Julian Schriftweiser, Demis Hassabis, Pushmeet Kohli, Martin Riedmiller, Oriol Vinyals, David Silver