CoT-based Synthesizer: Enhancing LLM Performance through Answer Synthesis

TL;DR

This paper introduces CoT-based Synthesizer, a novel inference scaling method that synthesizes better answers from multiple flawed candidates using Chain-of-Thought reasoning, improving large language models' accuracy efficiently.

Contribution

The paper presents a new inference scaling strategy leveraging Chain-of-Thought reasoning to synthesize superior answers, even with all candidates flawed, and introduces an automated data generation pipeline for training smaller models.

Findings

Significant accuracy improvements on benchmark datasets.

Enhanced performance for both small and large LLMs.

Open-source code and data for reproducibility.

Abstract

Current inference scaling methods, such as Self-consistency and Best-of-N, have proven effective in improving the accuracy of LLMs on complex reasoning tasks. However, these methods rely heavily on the quality of candidate responses and are unable to produce correct answers when all candidates are incorrect. In this paper, we propose a novel inference scaling strategy, CoT-based Synthesizer, which leverages CoT reasoning to synthesize superior answers by analyzing complementary information from multiple candidate responses, even when all candidate responses are flawed. To enable a lightweight and cost-effective implementation, we introduce an automated data generation pipeline that creates diverse training data. This allows smaller LLMs trained on this data to improve the inference accuracy of larger models, including API-based LLMs. Experimental results across four benchmark datasets with seven policy models demonstrate that our method significantly enhances performance, with gains of 11.8% for Llama3-8B and 10.3% for GPT-4o on the MATH dataset. The corresponding training data and code are publicly available on https://github.com/RUCKBReasoning/CoT-based-Synthesizer.