source: arxiv machine learning: from search to synthesis: training llms as zero-shot workflow generators
level: research
large language models often solve problems one instance at a time, which can lead to solutions that lack consistency and are hard to trust in real-world use. workflows that capture common patterns across many instances offer a more stable approach, providing clear steps for debugging and reuse. however, creating these workflows by hand takes deep knowledge and time, so automatic generation is desirable. current automatic methods either tailor solutions to single instances without learning broader patterns, or they cannot adapt to new tasks beyond their training setup.
metaflow treats workflow generation as a meta-learning challenge. the model learns to combine operators into strategies for a given task. training happens in two stages. first, it learns from a large collection of existing workflows. second, it refines its ability to generalize to unseen tasks. this two-stage process helps the model understand task-level structure rather than memorizing specific examples. the result is a system that can produce a workflow for a new task without any additional training, using only a description of the task and available operators.
the approach was tested on several domains, including data science pipelines and mathematical reasoning. metaflow generated workflows that matched or outperformed those designed by experts in terms of correctness and efficiency. the generated workflows were also easier to inspect and modify because they follow a clear, step-by-step format. this makes the system useful for automating repetitive analytical work while keeping human oversight possible. the method shows that with the right training, language models can move from answering one-off questions to building reusable solution templates.
why it matters: it enables reliable automation of complex analytical tasks by producing transparent, reusable workflows, reducing manual effort and errors in data science and ai pipelines.
source: arxiv machine learning: from search to synthesis: training llms as zero-shot workflow generators