Reproduce “FlowBank: Query-Adaptive Agentic Workflows Optimization through Precompute-and-Reuse” on OpenResearch

About this paper

Lingzhi Yuan, Chenghao Deng, Fangxu Yu, Souradip Chakraborty, Mohammad Rostami, Furong Huang

A framework for optimizing agentic workflows in large language model systems establishes a "precompute-and-reuse" paradigm, curating a compact bank of diverse workflows and adaptively selecting the most suitable one for each query. This approach achieved an average performance of 73.40 across five public benchmarks, outperforming state-of-the-art automated baselines by 3.00 percentage points while maintaining competitive inference costs.

Key insights

• Workflows deemed suboptimal and discarded during task-level optimization often hold substantial reuse value by solving different, complementary sets of queries.
• Query-level adaptivity can be largely recovered by adaptively selecting from a compact set of precomputed workflows, avoiding the high cost of per-query generation.
• Designing workflow search to prioritize complementarity, rather than just individual performance, is critical for building a diverse and broadly covering workflow bank.

Summary by alphaXiv.