According to a paper published on arXiv.org, researchers have introduced MetaFlow, a new approach that trains large language models to generate reusable workflows rather than instance-specific solutions. The method addresses a key limitation: while LLMs excel at individual tasks, their solutions often “lack the structural consistency needed for reliable deployment,” according to the paper.
MetaFlow treats workflow generation as a meta-learning problem, training models to compose solution strategies for a given task and operator set. The training occurs in two stages: supervised fine-tuning on synthetic workflow data, followed by reinforcement learning with verifiable rewards (RLVR) that uses execution feedback across problem instances.
According to the paper, MetaFlow achieves “performance comparable to state-of-the-art baselines on in-domain tasks with single inference, while demonstrating remarkable zero-shot generalization capabilities on out-of-domain tasks and operator sets.” The researchers tested the approach across benchmarks in question answering, code generation, and mathematical reasoning.
The paper notes that workflows “encode recurring algorithmic patterns at the task level,” offering robustness across instance variations, interpretable traces for debugging, and reusability. However, manually designing such workflows “requires significant expertise and effort,” which MetaFlow aims to address through automated generation. The 35-page paper includes 8 figures and is categorized under Machine Learning, Artificial Intelligence, and Computation and Language.