A research team wants language models to stop solving problems one at a time and start learning how to solve them in general.
The paper introduces MetaFlow, a system that reframes workflow generation as a meta-learning problem. Instead of producing a one-off answer to a question, MetaFlow trains a model to compose structured, reusable solution strategies from a set of available operators. Training happens in two stages: first, supervised fine-tuning on synthetic workflow data, then reinforcement learning using execution feedback across multiple problem instances. The idea is that the model learns the shape of a solution, not just the solution itself.
Most LLM deployments today are stateless guessers — each prompt gets its own answer, with no structural consistency between runs. Workflows that encode recurring algorithmic patterns offer something more useful: interpretable traces, easier debugging, and solutions that transfer across problem variations. MetaFlow's claim is that a model can learn to generate those workflows automatically, including for tasks and operator sets it was never trained on.
Across benchmarks in question answering, code generation, and mathematical reasoning, MetaFlow matches state-of-the-art baselines on familiar tasks and shows zero-shot generalization on new ones. Whether that holds outside tidy benchmark conditions is the question every paper like this leaves unanswered.