Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators

Dingkang Yang¹  Dongling Xiao²  Jinjie Wei¹  Mingcheng Li¹ 

Zhaoyu Chen¹  Ke Li³  Lihua Zhang¹ 

¹Fudan University

²ByteDance

³Tencent Youtu Lab

Abstract

Despite their remarkable capabilities, Large Language Models (LLMs) are prone to generate responses that contradict verifiable facts, i.e., unfaithful hallucination content. Existing efforts generally focus on optimizing model parameters or editing semantic representations, which compromise the internal factual knowledge of target LLMs. In addition, hallucinations typically exhibit multifaceted patterns in downstream tasks, limiting the model’s holistic performance across tasks. In this paper, we propose a Comparator-driven Decoding-Time (CDT) framework to alleviate the response hallucination. Firstly, we construct hallucinatory and truthful comparators with multi-task fine-tuning samples. In this case, we present an instruction prototype-guided mixture of experts strategy to enhance the ability of the corresponding comparators to capture different hallucination or truthfulness patterns in distinct task instructions. CDT constrains next-token predictions to factuality-robust distributions by contrasting the logit differences between the target LLMs and these comparators. Systematic experiments on multiple downstream tasks show that our framework can significantly improve the model performance and response factuality.

Deploy

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

conda create -n cdt python==3.8
conda activate cdt
pip install -r requirements.txt
cd ./transformers
pip install --editable ./

Datasets

The used datasets are organized as follows:

data
├── Alpaca_Gen
├── Alpaca_Judge
├── KNIGHT
├── truthfulqa
├── xsum

Models

• Please access here for the original llama models.
• The MoE-LoRA-based adaptors are used to collaborate with the base models to obtain the corresponding comparators.
• Please access here for testing BERTScore metrics.
• Please access here for testing DAE metrics.
• Please access here for testing FACTKB metrics.

llm_models
├── llama2-7b-base-cluster32-4moe-fact-adapter-hf # Truthful adaptor
├── llama2-7b-base-cluster32-4moe-halluc-adapter-hf # Hallucinatory adaptor
├── FactKB  # Test FactKB metrics
├── roberta-large  # Test BERTScore metrics

Run

For experiments on different datasets, please try:

cd ./exp_scripts/benchmark
sh ${dataset_name}.sh

Acknowledgement

We are aware that our works are inspired by the following works, including but not limited to

• ICD: https://github.com/HillZhang1999/ICD/tree/main
• LLaMA-Factory: https://github.com/hiyouga/LLaMA-Factory

Without these, nothing could happen in this repository.

If you are interested in our work, please cite:

@article{yang2024improving,
  title={Improving factuality in large language models via decoding-time hallucinatory and truthful comparators},
  author={Yang, Dingkang and Xiao, Dongling and Wei, Jinjie and Li, Mingcheng and Chen, Zhaoyu and Li, Ke and Zhang, Lihua},
  journal={arXiv preprint arXiv:2408.12325},
  year={2024}
}