Large Language Models Can Self-Improve

TL;DR

This paper demonstrates that large language models can self-improve their reasoning abilities by generating and fine-tuning on their own high-confidence answers using unlabeled data, achieving state-of-the-art results.

Contribution

The authors introduce a method for LLMs to self-improve through self-generated reasoning solutions without external labels, enhancing performance significantly.

Findings

Improved reasoning accuracy on multiple benchmarks.

Self-fine-tuning on generated solutions boosts model performance.

Self-improvement is effective even without ground truth labels.

Abstract

Large Language Models (LLMs) have achieved excellent performances in various tasks. However, fine-tuning an LLM requires extensive supervision. Human, on the other hand, may improve their reasoning abilities by self-thinking without external inputs. In this work, we demonstrate that an LLM is also capable of self-improving with only unlabeled datasets. We use a pre-trained LLM to generate "high-confidence" rationale-augmented answers for unlabeled questions using Chain-of-Thought prompting and self-consistency, and fine-tune the LLM using those self-generated solutions as target outputs. We show that our approach improves the general reasoning ability of a 540B-parameter LLM (74.4%->82.1% on GSM8K, 78.2%->83.0% on DROP, 90.0%->94.4% on OpenBookQA, and 63.4%->67.9% on ANLI-A3) and achieves state-of-the-art-level performance, without any ground truth label. We conduct ablation studies and show that fine-tuning on reasoning is critical for self-improvement.