A new framework called Workflow Optimisation (AWO) has been developed by researchers at the École Polytechnique Fédérale de Lausanne (EPFL) in Switzerland to improve the efficiency of complex workflows that utilize large language models (LLMs). The team, including Sami Abuzakuk, Anne-Marie Kermarrec, and Rishi Sharma, designed AWO to streamline processes by reducing redundant steps in workflows that depend heavily on iterative reasoning.
AWO analyzes workflow data to consolidate repeated sequences of tool calls into integrated meta-tools, which minimizes the reliance on LLM reasoning, helping to reduce operational costs and errors. The researchers reported a reduction in LLM calls by up to 11.9% and an increase in success rates by 4.2 percentage points. Their findings suggest that many workflows display consistent patterns, particularly in early stages, where over 14.3% of tasks follow similar sequences.
This regularity serves as the basis for AWO, which aims to enhance automation efficiency while maintaining flexibility. By meticulously analyzing existing workflow traces, AWO identifies common sequences, transforming them into reusable tools that enable a more reliable execution of tasks.