WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution
WorkRover claims gains by training cowork agents in environments that can repair their own tasks, context, and verifiers together.
The paper introduces recursive environment evolution for long-horizon agent training, aimed at keeping work context, executable tasks, and evidence-grounded verifiers aligned during refinement. It adds Factorial Compatibility Replay to test component revisions and choose the smallest validated repair. It also proposes Verifier-to-Turn Credit Assignment to give more granular reinforcement signals from workspace progress. A post-trained Qwen3.5-35B-A3B model is reported to beat the base model by 26.27 points on JobBench, 29.32 points on OfficeQA Pro, and 413.25 Elo on GDPval v2. HF Daily Papers' note
The paper introduces recursive environment evolution for long-horizon agent training, aimed at keeping work context, executable tasks, and evidence-grounded verifiers aligned during refinement. It adds Factorial Compatibility Replay to test component revisions and choose the smallest validated repair. It also proposes Verifier-to-Turn Credit Assignment to give more granular reinforcement signals from workspace progress. A post-trained Qwen3.5-35B-A3B model is reported to beat the base model by 26.27 points on JobBench, 29.32 points on OfficeQA Pro, and 413.25 Elo on GDPval v2. HF Daily Papers' note
score 5