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.
Introduction
Large Language Models (LLMs) excel at complex tasks when they break them down into smaller steps—a process known as task decomposition. However, current research often adopts a "one-size-fits-all" approach. Simple tasks might only need a basic Chain-of-Thought (CoT) approach, while complex multi-hop reasoning might require explicit planning and parallel execution. Using a complex method for a simple task wastes significant resources, while using a simple method for a complex task often leads to failure. This research addresses the fundamental performance-cost dilemma: explicit decomposition methods can be over 10x more expensive than implicit ones, yet no single method is optimal across all domains like math, coding, and creative writing.
Taxonomy of Task Decomposition
To understand the landscape of task decomposition, the authors conducted a massive empirical study across six categorization schemes. These include the sequence of decomposition and execution, the range of candidate subtasks, the form of model calls (implicit vs. explicit), the topology of subtask relationships (linear vs. DAG), the output format (text vs. code), and the use of external tools.
The Select-Then-Decompose (S&D) Strategy
Based on empirical insights, the researchers developed the Select-Then-Decompose (S&D) strategy. This framework consists of three collaborative modules designed to navigate the performance-cost trade-off:
- Selection Module: An LLM analyzes the input query's complexity and characteristics to select the most suitable decomposition approach (e.g., CoT, ReAct, or Plan-and-Execute).
- Execution Module: The system runs the selected approach to generate a candidate solution.
- Validation Module: A verifier LLM assesses the confidence of the result. If the confidence score falls below a threshold (e.g., 0.7), a staged switching mechanism triggers a more sophisticated (and usually more expensive) category of approaches until a reliable answer is found.
Empirical Insights and Principles
The study uncovered several critical principles for effective task decomposition:
- The Scaling Law of Execution: Scaling the execution model yields significantly higher gains than scaling the decomposition model. A smaller model can often handle planning, but a powerful reasoning model is required for execution.
- Task-Method Alignment: Different domains favor specific strategies. Math and logic benefit from implicit/linear strategies (CoT), creative writing favors DAG structures for parallel exploration, and coding benefits from interleaved strategies (ReAct) for iterative refinement.
- The Pareto Frontier: S&D aims to find the optimal balance by using low-cost strategies for the majority of tasks (~85%) and only escalating to expensive methods when necessary.
Experimental Evaluation
The framework was evaluated across diverse benchmarks including GSM8K/MATH (Reasoning), HumanEval (Coding), Trivia Creative Writing, and HotpotQA (Text Understanding).
Key findings include:
- Cost Efficiency: On HumanEval and HotpotQA, S&D achieved near-optimal performance while using only 24.77% of the average token cost of competitive methods.
- Pareto Optimality: Across all benchmarks, S&D consistently sat on the Pareto frontier, representing the best performance-to-cost ratio.
- Math Performance: On the difficult MATH benchmark, S&D achieved an accuracy of 52.39%, outperforming individual methods like CoT (50.73%) and ReAct (44.83%) while remaining significantly cheaper than pure Plan-and-Execute (P&E).
Conclusion
Select-Then-Decompose demonstrates that adaptive strategy selection is key to efficient LLM deployment. By dynamically matching task complexity with decomposition depth, S&D overcomes the performance-cost dilemma. Future work will explore automated prompt optimization for the selection module and the extension of this framework to multi-modal task decomposition.
Summary updated