LLMRefine: Pinpointing and Refining Large Language Models via Fine-Grained Actionable Feedback

TL;DR

LLMRefine is an inference-time optimization method that uses a learned feedback model to iteratively refine large language model outputs, significantly improving quality across multiple text generation tasks.

Contribution

This work introduces a novel inference-time refinement technique for LLMs using a learned feedback model and simulated annealing, enhancing output quality without additional human feedback.

Findings

Achieves up to 1.7 MetricX points improvement in translation

Improves ROUGE-L by 8.1 on long-form QA

Enhances ROUGE-L by 2.2 on topical summarization

Abstract

Recent large language models (LLM) are leveraging human feedback to improve their generation quality. However, human feedback is costly to obtain, especially during inference. In this work, we propose LLMRefine, an inference time optimization method to refine LLM's output. The core idea is to use a learned fine-grained feedback model to pinpoint defects and guide LLM to refine them iteratively. Using original LLM as a proposal of edits, LLMRefine searches for defect-less text via simulated annealing, trading off the exploration and exploitation. We conduct experiments on three text generation tasks, including machine translation, long-form question answering (QA), and topical summarization. LLMRefine consistently outperforms all baseline approaches, achieving improvements up to 1.7 MetricX points on translation tasks, 8.1 ROUGE-L on ASQA, 2.2 ROUGE-L on topical summarization.