Researchers Develop New Methods for Detecting LLM Hallucinations Without External References
Researchers have presented multiple approaches to detecting hallucinations in large language models without requiring external references, addressing a key challenge in deploying AI systems safely.
According to research published on arxiv.org and accepted at ICML 2026, a method called Human-like Criteria Probing for Hallucination Detection (HCPD) emulates “the multi-faceted reasoning of human evaluators” to detect when LLMs generate factually incorrect content. The approach uses a mechanism where an LLM agent “adaptively decomposes its judgment into a weighted set of interpretable criteria and aggregates criterion-specific scores into a final truthfulness measure,” relying solely on query-answer pairs without model internals or external references.
In a separate application, researchers from Atlassian developed HalluJudge for detecting hallucinations in LLM-generated code review comments, according to research accepted at FSE’26. The system achieved “an F1 score of 0.85 and an average cost of $0.009” across enterprise-scale software projects. According to the study, “67% of the HalluJudge assessments are aligned with the developer preference of the actual LLM-generated review comments in the online production.”
Additionally, researchers introduced Mental-R1, which applies aligned reasoning to mental health assessment. According to arxiv.org, the system uses Cognitive Relative Policy Optimization (CRPO) and achieved “an average improvement of 10.4 percentage points in weighted F1-score over the best reinforcement learning baseline” across 8 mental health datasets.