Unleashing the Second Brain: Enhancing Large Language Models through Chain of Thought with Human Feedback
Jia Jie YU, Minghui Luo, Honggu Zhou, Zhenzhong Lan · 2023
The expansion of large language models has led to improved performance and efficiency, with prompt engineering emerging as a key strategy for various LLM tasks. However, while these methods have proven beneficial, they are not without their constraints, particularly when it comes to sustaining a comprehensive, logical flow of ideas throughout the entire reasoning process. Despite the effectiveness of the Chain-of-Thought (CoT) and Tree-of-Thought (ToT) methods, they have limitations due to their end-to-end reasoning process. To address this, we introduce the chain-of-thought with human feedback. This new method incorporates human feedback into the model’s reasoning process, allowing for real-time adjustments and optimization. This approach fosters a more interactive and dynamic model operation, enabling the model to learn from human intuition and expertise, and improve its reasoning process over time. We have validated the effectiveness of our method through experiments in GSM8K and MMLU. Our main contributions include proposing the first human feedback to the chain of thought and the development of an intuitive interface for individuals to utilize large-scale models for problem-solving.