**Abstract:** SELF-DISCOVER is a framework designed to enable large language models (LLMs) to self-discover and compose task-specific reasoning structures, enhancing their performance on complex reasoning tasks. The core of the framework involves a self-discovery process where LLMs select and combine multiple atomic reasoning modules, such as critical thinking and step-by-step thinking, to form an explicit reasoning structure. This approach significantly improves the performance of LLMs like GPT-4 and PaLM 2 on challenging benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, achieving up to 32% improvement over Chain of Thought (CoT). SELF-DISCOVER also outperforms inference-intensive methods like CoT-Self-Consistency by more than 20% while requiring 10-40 times fewer inference computations. The discovered reasoning structures are found to be universally applicable across different model families and share commonalities with human reasoning patterns. **Introduction:** The paper introduces SELF-DISCOVER, a framework that enables LLMs to self-discover and compose task-specific reasoning structures, improving their performance on complex reasoning tasks. Unlike traditional prompting methods that rely on pre-defined modules, SELF-DISCOVER uses a meta-reasoning process to guide LLMs in selecting, adapting, and implementing relevant reasoning modules. This approach is inspired by how humans devise reasoning programs for problem-solving, involving three stages: selecting relevant modules, adapting them to the task, and implementing a structured plan. The framework is evaluated on diverse reasoning benchmarks, showing significant improvements over existing methods, particularly in tasks requiring world knowledge and algorithmic reasoning. Additionally, SELF-DISCOVER demonstrates efficiency in terms of inference calls, requiring only a few extra calls per instance compared to other methods. The paper also explores the universality of the discovered reasoning structures and their transferability between different LLMs, highlighting their potential for human-AI collaboration in complex problem-solving.**Abstract:** SELF-DISCOVER is a framework designed to enable large language models (LLMs) to self-discover and compose task-specific reasoning structures, enhancing their performance on complex reasoning tasks. The core of the framework involves a self-discovery process where LLMs select and combine multiple atomic reasoning modules, such as critical thinking and step-by-step thinking, to form an explicit reasoning structure. This approach significantly improves the performance of LLMs like GPT-4 and PaLM 2 on challenging benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, achieving up to 32% improvement over Chain of Thought (CoT). SELF-DISCOVER also outperforms inference-intensive methods like CoT-Self-Consistency by more than 20% while requiring 10-40 times fewer inference computations. The discovered reasoning structures are found to be universally applicable across different model families and share commonalities with human reasoning patterns. **Introduction:** The paper introduces SELF-DISCOVER, a framework that enables LLMs to self-discover and compose task-specific reasoning structures, improving their performance on complex reasoning tasks. Unlike traditional prompting methods that rely on pre-defined modules, SELF-DISCOVER uses a meta-reasoning process to guide LLMs in selecting, adapting, and implementing relevant reasoning modules. This approach is inspired by how humans devise reasoning programs for problem-solving, involving three stages: selecting relevant modules, adapting them to the task, and implementing a structured plan. The framework is evaluated on diverse reasoning benchmarks, showing significant improvements over existing methods, particularly in tasks requiring world knowledge and algorithmic reasoning. Additionally, SELF-DISCOVER demonstrates efficiency in terms of inference calls, requiring only a few extra calls per instance compared to other methods. The paper also explores the universality of the discovered reasoning structures and their transferability between different LLMs, highlighting their potential for human-AI collaboration in complex problem-solving.