FrontierSmith: Synthesizing Open-Ended Coding Problems Yields +309 Elo on ALE-Bench

arXiv·medium signal

Automated system evolves open-ended problems from competitive programming seeds by changing goals, restricting outputs, and generalizing inputs, using a quantitative idea divergence metric to select problems eliciting genuinely diverse solver approaches. Training Qwen3.5-27B on synthesized data yields +12.12 on FrontierCS and +309.12 Elo on ALE-bench; Qwen3.5-9B gains +8.82 and +306.36 respectively. Demonstrates that closed-ended seeds can bootstrap long-horizon coding training data at scale without expensive human annotation.