# Don't Let It Hallucinate: Premise Verification via Retrieval-Augmented Logical Reasoning

**Authors:** Yuehan Qin, Shawn Li, Yi Nian, Xinyan Velocity Yu, Yue Zhao, Xuezhe Ma  
**arXiv:** [2504.06438](https://arxiv.org/abs/2504.06438)  
**Date:** 2025-04-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2504.06438v1/

## Key points

- Authors address the issue of hallucinated outputs in large language models (LLMs) due to false premises in user queries
- Existing approaches rely on post-generation techniques that are computationally expensive and lack proactive mechanisms
- Proposed retrieval-based framework identifies and addresses false premises before generation using retrieval-augmented generation (RAG)
- Method involves transforming a user's query into a logical representation and assessing premise validity using factual sources
- Verification results are incorporated into the LLM's prompt to ensure factual consistency in the final output
- Experimental results show reduction in hallucinations, improved factual accuracy without requiring access to model logits or extensive fine-tuning
- Framework achieves high true positive rates, true negative rates, F1 scores, and overall accuracy by implementing logical forms for retrieval and original queries for false premise detection

## Summary

In their paper titled "Don't Let It Hallucinate: Premise Verification via Retrieval-Augmented Logical Reasoning," authors Yuehan Qin, Shawn Li, Yi Nian, Xinyan Velocity Yu, Yue Zhao, and Xuezhe Ma address the issue of hallucinated outputs in large language models (LLMs) when faced with false premises in user queries. These false premises can lead LLMs to generate fabricated or misleading information. Existing approaches to tackle this problem often rely on post-generation techniques that are computationally expensive and lack proactive mechanisms. To combat this issue, the authors propose a retrieval-based framework that identifies and addresses false premises before generation. The method involves transforming a user's query into a logical representation and using retrieval-augmented generation (RAG) to assess the validity of each premise using factual sources. The verification results are then incorporated into the LLM's prompt to ensure factual consistency in the final output. Experimental results demonstrate that this approach effectively reduces hallucinations, improves factual accuracy, and does not require access to model logits or extensive fine-tuning. By implementing logical forms for retrieval and original queries for false premise detection, the framework achieves high true positive rates, true negative rates, F1 scores, and overall accuracy. Overall,this innovative approach offers a promising solution to mitigating hallucinations in LLM-generated responses by proactively addressing false premises through logical reasoning and retrieval-based verification methods.