# Hallucination Detection in LLMs: Fast and Memory-Efficient Finetuned Models

**Authors:** Gabriel Y. Arteaga, Thomas B. Schön, Nicolas Pielawski  
**arXiv:** [2409.02976](https://arxiv.org/abs/2409.02976)  
**Date:** 2024-09-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.02976v1/

## Key points

- Researchers focus on distinguishing between faithfulness and factual hallucinations in Large Language Models (LLMs).
- Experiments evaluate performance by comparing proposed adaptations to Baseline models like BatchEnsemble with noise injection and prompt-based methods.
- Introduction of LoRA Ensemble approach for uncertainty-based experiments.
- Use of SQuAD and SQuAD 2.0 datasets for detecting faithfulness hallucinations by training LLMs to respond appropriately.
- Utilization of MMLU dataset for detecting factual hallucinations through multiple-choice question selection.
- Evaluation of predictive performance on downstream tasks using metrics like F1 score, exact match accuracy, and overall model accuracy.
- Conducting out-of-distribution tests by fine-tuning models on answerable questions from SQuAD 2.0 and evaluating them on unanswerable ones.
- Novel method presented for fast and memory-efficient training of LLM ensembles to detect both types of hallucinations effectively.
- Results show improved uncertainty estimates impacting model accuracy in high-risk AI implementation settings.

## Summary

In this study, the researchers focus on designing experiments to distinguish between faithfulness and factual hallucinations in Large Language Models (LLMs). They aim to evaluate the performance of their method for each type of hallucination. The experiments include comparing their proposed adaptations to Baseline models like BatchEnsemble with noise injection and prompt-based methods. Additionally, they introduce a LoRA Ensemble approach for uncertainty-based experiments. To detect faithfulness hallucinations, the researchers use the SQuAD and SQuAD 2.0 datasets, which consist of answerable and unanswerable questions. They train the LLMs to respond with "I don't know" for unanswerable questions and adjust training by including a balance of unanswerable questions to prevent hallucinations. For factual hallucination detection, they utilize the MMLU dataset, instructing models to select choices from multiple-choice questions. The study also evaluates predictive performance on downstream tasks such as SQuAD and MMLU datasets using metrics like F1 score, exact match accuracy, and overall model accuracy. Out-of-distribution tests are conducted by fine-tuning models on answerable questions from SQuAD 2.0 and evaluating them on unanswerable ones to assess their ability to recognize shifts in data distribution. Overall, this research presents a novel method for fast and memory-efficient training of LLM ensembles that can effectively detect both faithfulness and factual hallucinations. The results demonstrate improved uncertainty estimates that impact model accuracy in high-risk settings where AI implementation is crucial.