WorkRover: Fudan Unveils Recursive Environment Evolution for Cowork Agents, Surging JobBench by 26.27%

Background and the Problem

Scaling interactive environments is essential for training long-horizon cowork agents capable of cross-modal office automation, document analysis, and multi-turn workplace operations through reinforcement learning. However, current environment generation pipelines suffer from structural rigidity: synthetic environments remain static after initialization. When a defect arises in a task, rectifying it requires co-updating the surrounding work context and ground-truth verifiers. Existing frameworks lack mechanisms to maintain semantic alignment across evolving components, producing invalid rollouts or vulnerable verifiers prone to reward hacking.

Architecture and How It Works

To resolve this bottleneck, researchers from Fudan University NLP Lab introduced WorkRover, enabling recursive environment evolution:

1. Harness-Driven Synthesis: Autonomous construction agents jointly synthesize and validate semantically aligned work contexts, executable tasks, and evidence-grounded verifiers, adapting actions dynamically to workspace states and live execution outputs;
2. Factorial Compatibility Replay (FCR): Leverages rollout feedback to evaluate combinations of component revisions through dependency-closed replay, selecting the smallest validated repair among candidates while strictly preserving core constraints;
3. Verifier-to-Turn Credit Assignment (VTCA): Enhances trajectory-level GRPO advantages with deterministic workspace-progress checkpoints, tracing recorded tool-output and artifact dependencies to credit critical prerequisite turns.

Benchmarks and Measured Results

Post-training Qwen3.5-35B-A3B using WorkRover yielded WorkRover-35B-A3B, which established new empirical state-of-the-art results across complex workplace benchmarks:

• +26.27% Surge on JobBench: Achieves a 26.27 percentage-point gain over the base model on JobBench, excelling in realistic long-horizon job handovers;
• +29.32% Leap on OfficeQA Pro: Delivers a 29.32 percentage-point improvement on OfficeQA Pro, mastering complex spreadsheet and document reasoning;
• +413.25 Elo Surge on GDPval v2: Skyrockets by 413.25 Elo points on human-agent collaboration evaluations, demonstrating superior multi-turn coordination.

Getting Started for Developers

WorkRover provides an essential framework for organizations deploying enterprise-grade coworker agents:

• Paper Study: Review arXiv:2610.18067 for algorithmic details on dependency-closed replay and credit assignment mechanisms;
• Mitigate Sparse Reward Bottlenecks: Implement VTCA artifact dependency tracing to ensure intermediate milestones in long-horizon workflows receive meaningful policy gradients;
• Adaptive Sandbox Verification: Adopt FCR principles to keep simulated enterprise environments and unit tests automatically aligned with evolving business APIs.