FrontierSmith lands NeurIPS spotlight: AI-synthesized coding data beats expert curation

AccBalanced · x · 2026-09-25

The FrontierCS team announces FrontierSmith, a system for synthesizing open-ended coding training data at scale, has been accepted to NeurIPS as a spotlight. Starting from closed-ended tasks, it mutates, filters, and builds runnable optimization environments for long-horizon coding agents. Models trained on its data outperform human-expert-curated data on FrontierCS and ALE-bench, suggesting open-ended coding data is no longer the bottleneck. Paper, code, and model are open-sourced.

Original post →

More from coding & agent

coding & agent channel →