Contrastive Decoding Improves Reasoning in Large Language Models

TL;DR

Contrastive Decoding is a simple, training-free method that significantly enhances reasoning performance in large language models across various benchmarks by reducing errors and avoiding trivial copying modes.

Contribution

This paper demonstrates that Contrastive Decoding improves reasoning accuracy in large language models without additional training or complex modifications.

Findings

LLaMA-65B outperforms LLaMA 2, GPT-3.5, and PaLM 2-L on HellaSwag

LLaMA 65B surpasses LLaMA 2, GPT-3.5, and PaLM-540B on GSM8K

Contrastive Decoding outperforms nucleus sampling and greedy decoding

Abstract

We demonstrate that Contrastive Decoding -- a simple, computationally light, and training-free text generation method proposed by Li et al 2022 -- achieves large out-of-the-box improvements over greedy decoding on a variety of reasoning tasks. Originally shown to improve the perceived quality of long-form text generation, Contrastive Decoding searches for strings that maximize a weighted difference in likelihood between strong and weak models. We show that Contrastive Decoding leads LLaMA-65B to outperform LLaMA 2, GPT-3.5 and PaLM 2-L on the HellaSwag commonsense reasoning benchmark, and to outperform LLaMA 2, GPT-3.5 and PaLM-540B on the GSM8K math word reasoning benchmark, in addition to improvements on a collection of other tasks. Analysis suggests that Contrastive Decoding improves over existing methods by preventing some abstract reasoning errors, as well as by avoiding simpler modes such as copying sections of the input during chain-of-thought. Overall, Contrastive Decoding outperforms nucleus sampling for long-form generation and greedy decoding for reasoning tasks, making it a powerful general purpose method for generating text from language models.