MetaFlow: LLMs learn zero-shot workflow generation with RL
New training method replaces search with synthesis for reusable task workflows across domains.
A new paper from researchers (Luo et al.) introduces MetaFlow, a meta-learning framework that trains large language models to act as zero-shot workflow generators. Instead of producing instance-specific solutions, MetaFlow learns to compose solution strategies at the task level, encoding recurring algorithmic patterns into robust, interpretable, and reusable workflows. The training involves two stages: first, supervised fine-tuning on synthetic workflow data; second, reinforcement learning with verifiable rewards (RLVR) using execution feedback across multiple problem instances to improve end-to-end success.
MetaFlow is evaluated across question answering, code generation, and mathematical reasoning benchmarks. It achieves performance comparable to state-of-the-art baselines on in-domain tasks with a single inference pass, while demonstrating strong zero-shot generalization to out-of-domain tasks and novel operator sets. This approach addresses the bottleneck of manual workflow design by automating the synthesis of task-level patterns, potentially enabling more reliable and systematic deployment of LLMs in complex, multi-step problems.
- MetaFlow uses two-stage training: SFT on synthetic workflows then RL with verifiable rewards (RLVR).
- Achieves SOTA on in-domain QA, code, and math benchmarks with single inference.
- Demonstrates zero-shot generalization to unseen tasks and operator sets.
Why It Matters
Automates reliable, reusable workflow design for LLMs, enabling robust deployment across diverse domains.