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

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, extended

Summary

Human-like Criteria Probing for Hallucination Detection (HCPD) introduces an interpretable, zero-source method for identifying factually incorrect or unfaithful content generated by Large Language Models. Operating solely on query-answer pairs, HCPD employs an LLM agent that adaptively decomposes truthfulness judgments into a weighted set of interpretable criteria, such as factual accuracy and logical consistency, then aggregates criterion-specific scores. This adaptive capability is achieved through a reward-based alignment scheme utilizing weak supervision from semantic consistency metrics like BLEURT. At inference, a multi-sampling aggregation strategy ensures robust decisions. HCPD consistently outperforms state-of-the-art baselines, achieving an average AUROC of 88.19% on LLaMA-3.1-8b and 88.02% on Qwen-3-8b across datasets like TriviaQA, SciQ, NQ Open, and CoQA, demonstrating its effectiveness and explainability.

Key takeaway

For Machine Learning Engineers deploying LLMs in safety-critical applications or auditing black-box models, HCPD provides a robust and interpretable hallucination detection solution. Its zero-source, multi-criteria approach, validated with high AUROC scores, means you can assess model truthfulness without internal access or external knowledge. Consider integrating HCPD into your CI/CD pipeline for pre-deployment auditing or continuous monitoring to enhance trust and debug model outputs effectively.

Key insights

Zero-source LLM hallucination detection is enhanced by emulating human multi-criteria reasoning with adaptive, weakly-supervised agents.

Principles

• Decompose LLM evaluation into weighted, interpretable criteria.
• Align LLM evaluators using weak semantic consistency supervision.
• Multi-sampling aggregation stabilizes stochastic LLM judgments.

Method

An LLM agent adaptively generates context-aware criteria and weights, scores responses against them, and aggregates for a final truthfulness measure, trained via GRPO with weak semantic consistency supervision.

In practice

• Instantiate a Qwen-2.5-7b agent for zero-source detection.
• Apply multi-sampling (e.g., K=5) for robust hallucination scores.
• Utilize BLEURT or DeepSeek-V3 as weak supervision signals.

Topics

• LLM Hallucination Detection
• Zero-source Evaluation
• Human-like Criteria Probing
• Reward-based Alignment
• Group Relative Policy Optimization
• Explainable AI

Code references

• TRISKEL10N/HCPD
• huggingface/open-r1
• D2I-ai/eigenscore
• jlko/semantic_uncertainty
• collin-burns/discovering_latent_knowledge

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Zero-source LLM Hallucination Detection with Human-like Criteria Probing — HCPD detects LLM hallucinations in zero-source settings by emulating human multi-criteria reasoning for truthfulness, outperforming baselines.
• From Out-of-Distribution Detection to Hallucination Detection: A Geometric View — Reframing LLM hallucination detection as OOD detection offers a training-free, single-sample, scalable solution for reasoning tasks.
• Learning Uncertainty from Sequential Internal Dispersion in Large Language Models — SIVR detects LLM hallucinations by analyzing internal representation variance across layers and tokens.
• 5 Practical Techniques to Detect and Mitigate LLM Hallucinations Beyond Prompt Engineering — LLM hallucinations are a system problem requiring multi-layered detection and mitigation beyond prompt engineering.
• Weakly Supervised Distillation of Hallucination Signals into Transformer Representations — Hallucination signals can be distilled into LLM internal representations for detection without external inference-time resources.
• Quickest Detection of Hallucination Onset: Delay Bounds and Learned CUSUM Statistics — Sequential analysis and CUSUM-based methods are crucial for accurately measuring and minimizing hallucination detection delay in LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.