Self-Directed Decomposition Empowers Reasoning Potentials in Large Language Models

Chuan Tian, Yilei Zhang

Keywords: Large Language Models, Prompting Engineering, Self-Directed Decomposition

TL;DR: Self-Directed Decomposition enables LLMs to autonomously break down reasoning problems, boosting reasoning performance while revealing internal reasoning mechanisms.

Abstract: Large Language Models (LLMs) have demonstrated remarkable advancements in natural language processing and reasoning tasks, yet often struggle with logical coherence during problem-solving. This paper introduces Self-Directed Decomposition (SD), a novel prompting strategy enabling LLMs to autonomously decompose reasoning problems into manageable sub-tasks without human intervention, allowing models to determine their own approach with adaptive flexibility across diverse reasoning domains. Experiments across seven reasoning tasks reveal that this methodology particularly enhances performance on deductive, inductive, mathematical, commonsense, and scientific reasoning tasks, while showing more modest benefits for abductive and causal reasoning tasks, achieving 62.26\% overall median accuracy compared to 49.64\% and 46.43\% for zero-shot and zero-shot Chain-of-Thought (CoT) approaches, respectively. Error and statistical analysis demonstrates that SD significantly transforms reasoning patterns by reducing wrong selection errors but increasing process mistakes for simpler variants, with only SD1 maintaining optimal balance. We discover a counterintuitive negative correlation between token consumption and accuracy ($R^2 = 0.162, p = 0.004$), challenging conventional resource-performance assumptions. Abductive reasoning demonstrates critical vulnerability to decomposition strategies, showing significant perspective errors increase ($R^2 = 0.66$). These findings explain why SD1 outperforms other variants: it balances different error types effectively while avoiding the complexity-accuracy trade-off that affects simpler decomposition strategies.

Primary Area: Applications (e.g., vision, language, speech and audio, Creative AI)

Submission Number: 13998