DRACO adds 15.9 points to long-horizon agents

IBM researchers introduced DRACO, a training method that redistributes rubric judgments across responsible steps and improves AppWorld by 15.9 points over the base model and 5.3 points over sparse-reward GRPO, while gaining 5.3 points on out-of-domain Tau-Bench, according to the paper. It needs no trained attribution module or frontier judge, but the evidence covers two benchmarks and is preliminary; agent teams should file DRACO as an evaluation-design lead, not a production guarantee, alongside the harness study showing a five-task swing.