# Trapping LLM Hallucinations Using Tagged Context Prompts

**Authors:** Philip Feldman, James R. Foulds, Shimei Pan  
**arXiv:** [2306.06085](https://arxiv.org/abs/2306.06085)  
**Date:** 2023-06-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.06085v1/

## Key points

- Authors: Philip Feldman, James R. Foulds, Shimei Pan
- Challenge: Hallucinations in large language models (LLMs) like ChatGPT
- Importance of the issue as AI-driven platforms are widely adopted
- Proposed method: Recognize and flag instances of LLMs performing outside their domain knowledge
- Use of context and embedded tags to combat hallucinations in generative language models
- Baseline established for hallucination frequency without context
- Observation of significant reduction in overall hallucination when context is provided with question prompts
- Evaluation of the impact of placing tags within contexts on model responses
- Impressive 98.88% effectiveness achieved in eliminating hallucinations by employing this approach
- Contribution to addressing limitations of large language models by identifying and mitigating hallucinations
- Assurance of accurate information from AI-driven conversational agents like ChatGPT through incorporation of contextual information and embedded tags

## Summary

In their paper titled "Trapping LLM Hallucinations Using Tagged Context Prompts," authors Philip Feldman, James R. Foulds, and Shimei Pan address the challenge of hallucinations in large language models (LLMs) such as ChatGPT. While these sophisticated conversation agents have made significant advancements, they often generate false or fabricated information. This issue is particularly crucial as AI-driven platforms are being adopted across various sectors. To tackle this problem, the authors propose a novel method to recognize and flag instances when LLMs perform outside their domain knowledge, ensuring users receive accurate information. They demonstrate that combining context with embedded tags can effectively combat hallucinations within generative language models. The authors establish a baseline for hallucination frequency by examining prompt-response pairs without any context. They use generated URLs as indicators of fabricated data and observe a significant reduction in overall hallucination when context is provided along with question prompts for tested generative engines. Furthermore, the authors evaluate how placing tags within contexts impacts model responses. They achieve an impressive 98.88% effectiveness in eliminating hallucinations in responses by employing this approach. Overall, this research contributes to addressing the limitations of large language models by proposing a method to identify and mitigate hallucinations. By incorporating contextual information and embedded tags, users can be assured of receiving accurate information from AI-driven conversational agents like ChatGPT.