Select-Then-Decompose: From Empirical Analysis to Adaptive Selection Strategy for Task Decomposition in Large Language Models

TL;DR

This paper investigates task decomposition in large language models through a comprehensive taxonomy of decomposition approaches and an empirical analysis of factors that affect performance and cost. It introduces the Select-Then-Decompose framework, a closed-loop system with selection, execution, and verification modules that adaptively chooses decomposition approaches to balance performance and cost. Across multiple benchmarks, the approach consistently lies on the Pareto frontier, achieving strong results with substantially reduced token costs and improved reliability via verification. The work provides practical principles for method and model selection and releases open-source code to foster real-world deployment and further research.

Abstract

Large language models (LLMs) have demonstrated remarkable reasoning and planning capabilities, driving extensive research into task decomposition. Existing task decomposition methods focus primarily on memory, tool usage, and feedback mechanisms, achieving notable success in specific domains, but they often overlook the trade-off between performance and cost. In this study, we first conduct a comprehensive investigation on task decomposition, identifying six categorization schemes. Then, we perform an empirical analysis of three factors that influence the performance and cost of task decomposition: categories of approaches, characteristics of tasks, and configuration of decomposition and execution models, uncovering three critical insights and summarizing a set of practical principles. Building on this analysis, we propose the Select-Then-Decompose strategy, which establishes a closed-loop problem-solving process composed of three stages: selection, execution, and verification. This strategy dynamically selects the most suitable decomposition approach based on task characteristics and enhances the reliability of the results through a verification module. Comprehensive evaluations across multiple benchmarks show that the Select-Then-Decompose consistently lies on the Pareto frontier, demonstrating an optimal balance between performance and cost. Our code is publicly available at https://github.com/summervvind/Select-Then-Decompose.

Figures (9)

• Figure 1: Task decomposition approaches are categorized from six perspectives, with those highlighted in color representing the primary focus of this study, while the categories in gray are not within the scope of our investigation.
• Figure 2: Token cost and API call analysis across benchmarks. The bar represents the token cost, and the line represents the API calls.
• Figure 3: The left panel presents nine cross-model experiments, while the right panel shows three sets of controlled experiments.
• Figure 4: Performance comparison between reasoning and non-reasoning models in decomposition and execution stages.
• Figure 5: Performance vs. cost trade-offs across benchmarks