Synthetic Open-Ended Problem Generation
Introducing FrontierSmith, a new system that uses AI to synthesize open-ended coding problems at scale. High-quality open-ended coding tasks have long depended on expert human design, making them expensive and difficult to scale. FrontierSmith starts from closed-form algorithmic tasks and automatically mutates, filters, and constructs runnable optimization and testing environments, producing high-quality training tasks for long-horizon coding agents. In experiments, models trained on FrontierSmith-generated data outperform those trained on human-curated data on FrontierCS and ALE-bench.
Contributors
Runyuan He, Qiuyang Mang, Shang Zhou, Kaiyuan Liu, Hanchen Li, Huanzhi Mao, Qizheng Zhang, Zerui Li, Bo Peng, Lufeng Cheng, Tianfu Fu, Yichuan Wang, Wenhao Chai, Jingbo Shang, Alex Dimakis, Joseph E. Gonzalez, Alvin Cheung
Publications
CoRR – FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale