Back to feed
News Story
APriority78
量子位
1 sources

New Zero-Source LLM Hallucination Detection Method HCPD Achieves 88% Accuracy

A study accepted at ICML 2026 introduces HCPD, a method that detects hallucinations in large language models using only the question and answer text, without access to internal model information or external knowledge. By mimicking human-like dynamic evaluation, weak-supervised reward alignment, and multi-sampling aggregation, HCPD achieves over 88% AUROC on multiple datasets, significantly outperforming existing approaches.

SynthePulse Insight · AI deep readingMembers

New Breakthrough in Zero-Source Hallucination Detection: How HCPD Achieves 88% Accuracy with Human-Like Criteria Probing

Version 1 · 1 source

In extreme scenarios where access to model internals is unavailable and external knowledge bases are absent, how can we determine whether a large language model's response is trustworthy? At ICML 2026, HCPD introduces a human-like dynamic evaluation mechanism that adaptively generates evaluation criteria and aggregates multiple assessments, achieving significantly better detection performance than existing methods across multiple datasets.

  • HCPD targets the Zero-source hallucination detection setting: the detector can only rely on the question and answer itself to judge reliability, with no access to the target model's output probabilities, intermediate states, or other internal information, and it is difficult to rely on search engines, knowledge bases, or reference answers.
  • The core mechanism is human-like criteria probing: the agent starts from general dimensions such as factual correctness and logical consistency, adaptively generates fine-grained evaluation criteria for specific question-answer pairs, assigns weights, then scores each dimension and makes a comprehensive judgment.
  • The training phase uses GRPO for reward alignment, constructing weak supervision labels via semantic similarity, mapping answer reliability to a fine-grained 1-10 scoring scale to provide more discriminative optimization signals.
Open section navigationZero-Source Setting: Hallucination Detection Without Reference Answers or Internal Information

Zero-Source Setting: Hallucination Detection Without Reference Answers or Internal Information

Large language models are widely used in decision-support scenarios such as education and healthcare, but hallucination remains a core risk to their reliable deployment. Existing detection methods typically rely on auxiliary signals such as the target model's output probabilities, confidence scores, multiple generation results, or internal representations. However, in real-world applications, the source of the text to be detected is often unknown, and the detector may only have access to the final question-answer text. Continuously relying on knowledge bases or reference answers for fact-checking not only incurs additional computational and time costs, but its effectiveness is also limited by the coverage and reliability of external information.

To address this, the HCPD team proposes a stricter setting—Zero-source hallucination detection: the detector can only judge reliability based on the given question and answer itself. This significantly compresses the available information and is closer to some real deployment environments. The paper was published at ICML 2026, titled "Zero-source LLM Hallucination Detection with Human-like Criteria Probing."

Free for now

Read the full analysis

4 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

The information in this article primarily comes from a report by QbitAI on the ICML 2026 paper, which is a secondary source. Links to the paper and code are provided, but the experimental data have not been independently verified. All performance figures (such as 88.19% and 89.62%) are as reported in the paper and should be considered source claims rather than confirmed facts.

Sources for this version

  1. 只靠一问一答，就能抓出大模型幻觉，准确率88% | ICML'26量子位

Primary report

量子位

Primary source