LLMs Trained as Zero-Shot Workflow Generators for Task Automation

Researchers have developed MetaFlow, a novel approach that transforms Large Language Models (LLMs) into zero-shot workflow generators. Unlike existing methods that produce instance-specific solutions or struggle with generalization, MetaFlow aims to create structured, reusable algorithmic patterns at the task level. This provides greater robustness, interpretability for debugging, and reusability across different problem instances. The MetaFlow training process involves two key stages. Initially, the model undergoes supervised fine-tuning using synthetically generated workflow data. This is followed by a reinforcement learning phase, termed RLVR (Reinforcement Learning with Verifiable Rewards), which leverages execution feedback from various problem instances to enhance the model's overall success rate. Evaluations across diverse benchmarks, including question answering, code generation, and mathematical reasoning, demonstrate MetaFlow's effectiveness. It achieves performance comparable to state-of-the-art baselines on tasks it was trained on, and critically, exhibits strong generalization capabilities to entirely new, untrained tasks and novel sets of operators.