Zero-source LLM Hallucination Detection with Human-like Criteria Probing

2026-06-11

AI Summary

This work addresses the challenge of evaluating the factual consistency of large language model (LLM) generations in a zero-source setting—where neither internal model information nor external references are available. The authors propose the Human-Centric Probing (HCP) mechanism, which mimics human evaluators’ multidimensional reasoning by decomposing truthfulness judgments into interpretable criteria and aggregating them with learned weights, using only question-answer pairs to detect hallucinations. Innovatively, they introduce a weakly supervised semantic consistency–based reward alignment framework that enables adaptive, interpretable, and external-knowledge-free hallucination detection. By integrating LLM agents, human-like probing, multi-sample aggregation, and semantic consistency alignment, the method significantly outperforms existing approaches across multiple benchmarks, achieving efficient, robust, and explainable zero-source hallucination detection.

Abstract

Large language models (LLMs) often hallucinate by generating factually incorrect or unfaithful content, posing significant risks to their safe use. Detecting such hallucinations is particularly challenging under the zero-source constraint, where no model internals or external references are available, and detection must rely solely on the textual query-answer pair. In this paper, we propose Human-like Criteria Probing for Hallucination Detection (HCPD), a paradigm that emulates the multi-faceted reasoning of human evaluators. Its core is a Human-like Criteria Probing (HCP) mechanism, in which a LLM agent adaptively decomposes its judgment into a weighted set of interpretable criteria and aggregates criterion-specific scores into a final truthfulness measure. To achieve this adaptive capability, we introduce a reward-based alignment scheme using only weak supervision from semantic consistency. At inference, we employ a multi-sampling aggregation strategy to ensure robust decisions while preserving full interpretability. We further provide theoretical analysis supporting the reliability of our approach. Extensive experiments show that HCPD consistently outperforms state-of-the-art baselines, offering an effective and explainable solution for zero-source hallucination detection. Code is available at https://github.com/TRISKEL10N/HCPD.

Problem

• LLM hallucination
• zero-source detection
• truthfulness evaluation
• factuality assessment
• query-answer pair

Innovation

• zero-source hallucination detection
• human-like criteria probing
• interpretable evaluation
• reward-based alignment
• multi-sampling aggregation