Self-Discover: Large Language Models Self-Compose Reasoning Structures (AI summary)

Key Points

1. The paper introduces S ELF -D ISCOVER, a framework for Large Language Models (LLMs) to self-discover task-intrinsic reasoning structures to efficiently tackle complex reasoning problems.

2. S ELF -D ISCOVER substantially improves LLMs' performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by as much as 32% compared to Chain of Thought (CoT), while requiring 10-40x fewer inference compute.

3. The framework operates in two stages: in Stage 1, the LLM selects, adapts, and implements reasoning structures for a specific task, and in Stage 2, the LLM follows the self-discovered structure to arrive at the final answer during decoding.

4. S ELF -D ISCOVER is efficient in computation as it only requires 3 more inference steps on the task-level, while outperforming inference-intensive methods such as CoT-Self-Consistency and majority voting of every module.

5. The self-discovered reasoning structures are universally applicable across model families and share commonalities with human reasoning patterns, demonstrating their effectiveness and transferability between different LLMs.

6. The paper presents empirical evidence that S ELF -D ISCOVER outperforms other zero-shot prompting methods for LLM reasoning and specifically highlights its superiority in tasks requiring world knowledge.

7. Ablation studies show that all three actions involved in the self-discovery process (SELECT, ADAPT, IMPLEMENT) are beneficial for task-solving and the reasoning structures are adapted to be task-specific, providing the most gain in solving reasoning tasks.

8. The study demonstrates the universality of the self-discovered reasoning structures by applying the structures discovered by one LLM to another, showing robust performance across different LLMs.

9. The paper concludes by emphasizing S ELF -D ISCOVER as a potent framework for models to self-discover reasoning structures, highlighting the potential for Human-AI collaboration and pushing the boundaries of problem-solving using LLMs.

Summary

Introduction of SELF-DISCOVER framework
The paper introduces a framework called SELF-DISCOVER, which enables Language Model Learners (LLMs) to autonomously uncover task-intrinsic reasoning structures. The self-discovery process involves LLMs choosing multiple atomic reasoning modules and composing them into an explicit reasoning structure. The study demonstrates substantial performance improvements of SELF-DISCOVER compared to existing methods, particularly on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH. The self-discovered reasoning structures are found to be universally applicable across different model families and align with human reasoning patterns.

Improvement of reasoning capabilities in LLMs
The research focuses on improving the reasoning capabilities of LLMs and explores the process of self-discovering the underlying reasoning structure unique to each task. The study introduces SELF-DISCOVER as a framework for LLMs to self-discover task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. The core of the framework is a self-discovery process where LLMs select multiple atomic reasoning modules and compose them into an explicit reasoning structure for decoding. The research shows substantial performance improvements of SELF-DISCOVER on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, compared to existing methods like Chain of Thought (CoT).

Furthermore, SELF-DISCOVER outperforms inference-intensive methods such as CoT-Self-Consistency by more than 20%, while requiring 10-40x fewer inference computations. The self-discovered reasoning structures are found to be universally applicable across different model families, including GPT-4 and PaLM 2. The paper also discusses the importance of using the reasoning structures discovered through the self-discovery process and demonstrates the efficiency and effectiveness of the proposed framework.

Reference: https://arxiv.org/abs/2402.03620