Carnegie Mellon's Hybrid-Gym trains coding agents with 25% better generalization

Researchers from Carnegie Mellon University and Meta built Hybrid-Gym, a training environment for coding agents. It uses scalable synthetic tasks like function localization to teach transferable skills. Agents trained with it showed a 25.4% absolute gain on SWE-Bench Verified and improved on other benchmarks. This approach helps AI models generalize to complex, real-world coding tasks they haven't seen before, moving beyond single-issue fixes.