"Sorry, Come Again?" Prompting - Enhancing Comprehension and Avoiding Hallucination with [PAUSE]-injected Optimal Paraphrasing

ACL ARR 2025 February Submission1510 Authors

Abstract: Hallucination has emerged as the most vulnerable aspect of Large Language Models (LLMs). This paper introduces Sorry, Come Again (SCA) prompting to avoid hallucinations by improving comprehension through optimal paraphrasing and injecting [PAUSE] tokens to delay LLM generation. We analyze the linguistic nuances - formality, readability, and concreteness - of prompts for 22 LLMs and their impact on hallucinations. The lack of these nuances makes it harder for LLMs to understand prompts, leading them to generate speculative content based on memory, which can be inaccurate. We also explore the phenomenon of ``lost in the middle,'' where LLMs neglect the middle sections of prompts. To address this, we propose an optimal paraphrasing technique and evaluate it using Integrated Gradients to ensure accurate processing. Additionally, we inject [PAUSE] tokens to help LLMs better comprehend longer prompts by mimicking human reading pauses, optimizing their placement and number. We introduce reverse knowledge distillation to fine-tune the model for better [PAUSE] insertion. Finally, we introduce ACTIVATOR, an end-to-end framework that enhances LLMs' reading comprehension to avoid hallucinations. The SCA demo is publicly available at \href{https://huggingface.co/spaces/aisafe/SCA}{link}.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: feature attribution,probing,robustness,topic modeling

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 1510