DR-CoT: dynamic recursive chain of thought with meta reasoning for parameter efficient models

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Chain-of-Thought (CoT) prompting enhances Large Language Model (LLM) reasoning but faces challenges with high computational costs and context dilution.
• These limitations hinder LLM effectiveness in resource-constrained and real-time applications.

Purpose of the Study:

• Introduce Dynamic Recursive Chain-of-Thought (DR-CoT), a novel framework designed to overcome the limitations of traditional CoT prompting.
• Enhance reasoning accuracy and efficiency in parameter-efficient models through a synergistic approach.

Main Methods:

• DR-CoT integrates recursive reasoning, dynamic context truncation, and a voting mechanism to manage context within a fixed token budget.
• Multiple independent reasoning chains are aggregated to improve inference and accuracy.

Main Results:

• DR-CoT achieved notable accuracy gains on challenging benchmarks like GPQA Diamond (1.5%-4.4%) and AIME2024 (3-4 percentage points).
• Improved zero-shot classification performance, enabling smaller models to outperform larger ones like GPT-4.
• Outperformed established frontier LLMs in code generation tasks on HumanEval.

Conclusions:

• DR-CoT effectively bridges the performance gap between parameter-efficient models and state-of-the-art LLMs across diverse domains.
• The framework offers a computationally efficient and accurate solution for complex reasoning tasks.