AlphaDev
RL agent playing an assembly-instruction game found sorting routines up to 70% faster on short inputs, merged into LLVM libc++.
Frames algorithm discovery as a single-player assembly game, searching at instruction level rather than rewriting human algorithms. First RL-discovered algorithm added to libc++ sorting; hashing routine for 9-16 bytes 30% faster, added to Abseil.
- Date
- Wednesday, 7 June 2023
- Lab
- Google DeepMind
- Kind
- paper
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Sort speedup (short sequences) | up to 70% 1.7% for sequences over 250,000 elements | company |
| Hashing speedup (9-16 bytes) | ~30% adopted into Abseil | company |
Published in Nature 2023-06-07. The discovered routines are in public libraries (LLVM libc++, Abseil); the agent itself was not released.
Sources
This record was checked against its sources on 6 October 2026. How we check