<div align="center">

LoGU: Long-form Generation with Uncertainty Expressions

<div> <a href='https://scholar.google.com/citations?user=asTSVwQAAAAJ&hl=en' target='_blank'><b>Ruihan Yang</b></a>¹&emsp; <a href='https://caiqizh.github.io/' target='_blank'><b>Caiqi Zhang</b></a>²&emsp; <a href='https://scholar.google.co.jp/citations?user=373vlUEAAAAJ&hl=en' target='_blank'><b>Zhisong Zhang</b></a>³&emsp; </div> <div>¹Fudan University</div> <div>²University of Cambridge</div> <div>³Tencent AI Lab</div> <div> <h4>

<img src="https://img.shields.io/badge/Version-1.0-blue.svg" alt="Version"> <img src="https://img.shields.io/github/stars/rhyang2021/LoGU?color=yellow" alt="Stars"> <img src="https://img.shields.io/github/issues/rhyang2021/LoGU?color=red" alt="Issues">

</h4> </div> <img width="405" alt="image" src="./figures/head_png.png" style="display: block; margin: 0 auto;"> </div> <!-- 关闭外层的居中div -->

Introduction

While Large Language Models (LLMs) demonstrate impressive capabilities, they still struggle with generating factually incorrect content (i.e., hallucinations). A promising approach to mitigate this issue is enabling models to express uncertainty when unsure. Previous research on uncertainty modeling has primarily focused on short-form QA, but real-world applications often require much longer responses. In this work, we introduce the task of Long-form Generation with Uncertainty (LoGU). We identify two key challenges: Uncertainty Suppression, where models hesitate to express uncertainty, and Uncertainty Misalignment, where models convey uncertainty inaccurately.

To tackle these challenges, we propose a refinement-based data collection framework and a two-stage training pipeline. Our framework adopts a divide-and-conquer strategy, refining uncertainty based on atomic claims. The collected data are then used in training through supervised fine-tuning (SFT) and direct preference optimization (DPO) to enhance uncertainty expression. Extensive experiments on three long-form instruction following datasets show that our method significantly improves accuracy, reduces hallucinations, and maintains the comprehensiveness of responses.

<div align="center"> <img width="825" alt="image" src="./figures/main_png.png"> </div>

How to Install

You can use the following commands to install the environment for LoGU:

conda create -n LoGU python==3.8
conda activate LoGU
pip install -r lf_requirements.txt
pip install -r vllm_requirements.txt

Run

Try the following command to test our method on Bios, LongFact, WildHallu:

• Generate answers

cd ./scripts
bash generate_vllm_responses.sh

• Calculate Factual Accuracy(FA)

bash eval_pipeline.sh

• Calculate Uncertain Precision(UC)

bash generate_unc_answers.sh
bash factcheck_unc_answers.sh

Training Data

Training data for LoGU-SFT and LoGU-DPO in the paper can be found here.

Models

We also provide some uncertainty expression models on the huggingface model hub for fast trail:

| Model | Link | | :------- | :---------: | | rhyang2021/uncertain_llama3_8b | HuggingFace| | rhyang2021/uncertain_mistral_7b | HuggingFace|

If you have any questions, please feel free to email me or drop me an issue.