Think Twice Before Trusting: Self-Detection for Large Language Models through Comprehensive Answer Reflection
Moxin Li, Wenjie Wang, Fuli Feng, Fengbin Zhu, Qifan Wang, Tat‐Seng Chua · 2024
Self-detection for Large Language Models (LLMs) seeks to evaluate the trustworthiness of the LLM's output by leveraging its own capabilities, thereby alleviating the issue of output hallucination.However, existing self-detection approaches only retrospectively evaluate answers generated by LLM, typically leading to the over-trust in incorrectly generated answers.To tackle this limitation, we propose a novel selfdetection paradigm that considers the comprehensive answer space beyond LLM-generated answers.It thoroughly compares the trustworthiness of multiple candidate answers to mitigate the over-trust in LLM-generated incorrect answers.Building upon this paradigm, we introduce a two-step framework, which firstly instructs LLM to reflect and provide justifications for each candidate answer, and then aggregates the justifications for comprehensive target answer evaluation.This framework can be seamlessly integrated with existing approaches for superior self-detection.Extensive experiments on six datasets spanning three tasks demonstrate the effectiveness of the proposed framework. * * Corresponding author.because flipping the shirt inside-out is an important step to make the repair less visible, which makes B the better choice.Read the given question and select the most appropriate answer.How do you repair a torn shirt? A. Prepare the needle and thread.Pull together the fabric and sew together.B. Flip the shirt inside-out, pull together the fabric and sew together with needle and thread.