Think Twice Before Trusting: Self-Detection for Large Language Models through Comprehensive Answer Reflection

TL;DR

This paper introduces a new self-detection framework for LLMs that evaluates multiple candidate answers through reflection and justification, reducing over-trust in incorrect outputs and improving trustworthiness assessment.

Contribution

It proposes a comprehensive answer reflection paradigm and a two-step framework that enhances self-detection by considering multiple answers and their justifications.

Findings

Effective in reducing over-trust in incorrect answers

Improves self-detection accuracy across multiple datasets

Seamlessly integrates with existing LLM self-evaluation methods

Abstract

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 self-detection 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.