Self-Harmonized Chain-of-Thought (ECHO) is an advanced technique that enhances Chain-of-Thought (CoT) prompting in Large Language Models (LLMs) by refining multiple reasoning paths into a unified pattern.
Traditional Chain-of-Thought (CoT) prompting allows LLMs to break down complex problems into intermediate steps, either by using simple prompts like “Let’s think step by step” (Zero-Shot-CoT) or with human-crafted examples (Few-Shot-CoT). ECHO builds on this by improving how LLMs handle diverse solution paths, using an iterative process to harmonize these variations into a consistent and more accurate reasoning approach.
ECHO improves on traditional CoT methods by addressing two key limitations:
ECHO’s key innovation is its dynamic, self-harmonization process, where demonstrations are continuously refined through multiple iterations. The method involves:
This harmonization reduces errors and aligns different reasoning paths into a coherent framework.
ECHO can be applied to a wide range of reasoning tasks, including arithmetic, commonsense, and symbolic reasoning. Here’s a simple template for how you might use it in an AI system:
Clustering questions based on similarity.
Generating rationales for each question using Zero-Shot-CoT prompts.
[Question from step 1]
Let's think step by step.
ECHO was tested on three major reasoning domains: arithmetic, commonsense, and symbolic reasoning. Below are the performance improvements ECHO achieved compared to other methods:
Method | Arithmetic | Commonsense | Symbolic | Overall |
---|---|---|---|---|
Zero-Shot-CoT | 77.3% | 61.4% | 63.1% | 71.3% |
Few-Shot-CoT | 82.1% | 69.7% | 88.5% | 80.9% |
Auto-CoT | 80.8% | 65.7% | 87.8% | 79.2% |
ECHO | 83.1% | 70.5% | 90.3% | 82.0% |
ECHO demonstrates the best overall performance, especially in symbolic reasoning, where it outperforms all other methods. Its harmonized approach makes it more effective in generating consistent and correct reasoning across various problem types.
Sander Schulhoff is the Founder of Learn Prompting and an ML Researcher at the University of Maryland. He created the first open-source Prompt Engineering guide, reaching 3M+ people and teaching them to use tools like ChatGPT. Sander also led a team behind Prompt Report, the most comprehensive study of prompting ever done, co-authored with researchers from the University of Maryland, OpenAI, Microsoft, Google, Princeton, Stanford, and other leading institutions. This 76-page survey analyzed 1,500+ academic papers and covered 200+ prompting techniques.
Jin, Z., & Lu, W. (2024). Self-Harmonized Chain of Thought. https://arxiv.org/abs/2409.04057 ↩