We Trained a Model on Office Work. It Also Got Better at Coding (surgehq.ai)

Researchers trained the Qwen3.5-122B-A10B model using a variety of reinforcement learning (RL) tasks related to office work, such as document handling and web research, but surprisingly found significant improvements in the model's coding capabilities as well. Despite the absence of direct coding tasks in the training process, the model demonstrated a remarkable 5.8 percentage point increase in software engineering benchmarks it hadn't encountered before. This raises intriguing questions about the transferability of skills, as the model learned to effectively form and pursue goals, maintain situational awareness, and manage complex workflows—all essential for successful coding tasks. The training highlighted the importance of hierarchical goal-directed execution. Each task involved breaking down high-level objectives into smaller, manageable subgoals, allowing the model to build a coherent understanding of its environment and objectives over multiple iterations. These capabilities proved crucial when handling dynamic and interconnected information, showcasing that the skills acquired through office tasks could successfully inform coding practices. This unexpected cross-domain improvement suggests that foundational goal-management skills may be more broadly applicable than previously understood, offering valuable insights for future AI/ML training methodologies.