The Tyranny of Possibilities in the Design of Task-Oriented LLM Systems: A Scoping Survey

TL;DR

This survey explores the design space of task-oriented LLM systems, proposing conjectures and a new agent-centric perspective to guide future research in system configurations, prompting techniques, and efficiency considerations.

Contribution

It introduces seven conjectures about LLM system design, proposes an agent-centric framework for prompting, and highlights gaps in efficiency evaluation, guiding future research directions.

Findings

Identified a pattern in LLM system configurations and hypothesized outcomes.

Proposed an agent-centric perspective on prompting techniques.

Highlighted the lack of focus on computational and energy efficiency.

Abstract

This scoping survey focuses on our current understanding of the design space for task-oriented LLM systems and elaborates on definitions and relationships among the available design parameters. The paper begins by defining a minimal task-oriented LLM system and exploring the design space of such systems through a thought experiment contemplating the performance of diverse LLM system configurations (involving single LLMs, single LLM-based agents, and multiple LLM-based agent systems) on a complex software development task and hypothesizes the results. We discuss a pattern in our results and formulate them into three conjectures. While these conjectures may be partly based on faulty assumptions, they provide a starting point for future research. The paper then surveys a select few design parameters: covering and organizing research in LLM augmentation, prompting techniques, and uncertainty estimation, and discussing their significance. The paper notes the lack of focus on computational and energy efficiency in evaluating research in these areas. Our survey findings provide a basis for developing the concept of linear and non-linear contexts, which we define and use to enable an agent-centric projection of prompting techniques providing a lens through which prompting techniques can be viewed as multi-agent systems. The paper discusses the implications of this lens, for the cross-pollination of research between LLM prompting and LLM-based multi-agent systems; and also, for the generation of synthetic training data based on existing prompting techniques in research. In all, the scoping survey presents seven conjectures that can help guide future research efforts.