Using LLMs for complex tasks is challenging "involving a time-consuming and uncontrollable prompt engineering process", hence this "human-LLM interaction framework, Low-code LLM... six types of simple low-code visual programming interactions. ... [Via] a graphical user interface, users can incorporate their ideas into the workflow without writing trivial prompts...:
- Planning LLM designs a structured planning workflow for complex tasks... edited and confirmed by users through low-code visual programming
- Executing LLM generates responses following the user-confirmed workflow"
The paper, from a team at Microsoft, sets out three advantages of the framework, "soon publicly available at LowCodeLLM":
- controllable generation of results... Complicated tasks are decomposed into structured conducting plans and presented to users as workflows... responses ... more aligned with user’s requirements
- user-friendly human-LLM interaction: ... swiftly comprehend the LLMs’ execution logic ... [GUI] empowers users to conveniently modify the workflow
- wide applicability
Four task categories:
- "Long Content generation, including long texts ... and posters...
- Large Project Development, including complex object relations and system design...
- Task-completion Virtual assistant, where developers predefine the interaction logic between virtual assistant and customers...
- Knowledge-embedded system, where domain experts can embed their experience or knowledge into a conducting workflow. Then, the counseling assistant will follow a pre-defined pattern and act as a coach to scaffold users to complete their tasks.