Factors and trade-off in LLM task decomposition

Determine the key factors that influence both performance and token/computational cost in task decomposition approaches for large language models, and ascertain how to balance the trade-off between performance and cost across tasks and settings.

Background

The paper studies task decomposition in LLMs and notes that prior work has focused on enhancing performance via tools, feedback, and memory while often overlooking the cost-performance trade-off. In the introduction, the authors explicitly flag unresolved questions regarding which factors drive performance and cost and how to balance them.

These questions motivate the empirical analyses and the proposed Select-Then-Decompose strategy. Although the paper presents insights and a strategy, the introduction frames the identification of factors and principles for balancing performance against cost as an explicitly unanswered question in the field, warranting inclusion as an open problem.