Hallucination Detector Arena
Can we trust LLM hallucination detectors? Four detectors, measured on 480 answers with known ground truth across six domains, plus a live playground. By Mohammed Achraf Larbi.
- Best at ranking answers: MiniCheck (RoBERTa-Large), test AUROC 0.67 (95% CI 0.60–0.74).
- Performance depends on the domain: MiniCheck goes from AUROC 0.92 on CovidQA (biomedical) to 0.49 on DROP (reading comprehension) (40 test answers per domain, so per-domain values are uncertain).
- Hard domains: no detector reaches AUROC 0.65 on DROP (reading comprehension), FinanceBench (finance), PubMedQA (biomedical): close to chance (0.5).
- Detectors often disagree: pairwise Cohen's κ from 0.12 to 0.50 (1 = perfect agreement); swapping detectors changes which answers get flagged.
- Blind spots: 31 of 240 test examples fool every detector.
LLM judge (gpt-oss-120b): evaluation in progress (224/480 answers scored, limited by the LLM API's free daily token quota). It joins the comparison once it has scored every answer; meanwhile it is available live in Try it.
Leaderboard (test half)
NLI baseline (DeBERTa-v3-base) | 0.672 (0.60–0.74) | 0.592 | 0.08 | 0.613 | 0.583 | 2.74 s |
Do detectors agree? (test half, tuned thresholds)
NLI baseline | Vectara HHEM-2.1-Open | 0.12 | 0.566 | 42 | 240 |
Paste a context, a question and an answer: each detector says whether the answer is supported. Verdicts use the thresholds tuned in the study.
Question
Hallucination detectors are increasingly used to audit LLM answers, but are the detectors themselves valid? A detector you cannot trust turns every downstream "hallucination rate" into noise. This project measures four detectors on answers whose truth is known, before using them for anything else.
Data
HaluBench (Patronus AI, CC-BY-NC-2.0): context + question + answer triples labelled faithful or hallucinated, from six sources: HaluEval, DROP, CovidQA, PubMedQA, FinanceBench and RAGTruth. The study uses a balanced sample of 480 examples (40 faithful + 40 hallucinated per source, contexts of at most 1,500 words), split 50/50 into dev and test halves per source and label.
Protocol
- Every detector outputs a support score in [0, 1]; low support = hallucination.
- AUROC (threshold-free ranking quality) is reported on the test half with a 95% bootstrap CI (2,000 resamples).
- Decision thresholds are tuned on dev only (maximising balanced accuracy) and evaluated on test, next to the default 0.5, so the reported accuracy is not inflated by choosing the threshold on the evaluation data.
- Short answers ("0.25", "Rams") cannot be checked alone, so each claim is question + answer.
- Pairwise McNemar tests and Cohen's κ compare detectors on the same test answers.
Detectors
| Detector | Type | License |
|---|---|---|
| NLI baseline | cross-encoder/nli-deberta-v3-base, P(entailment), max over context chunks |
Apache-2.0 |
| MiniCheck | lytang/MiniCheck-RoBERTa-Large (Tang et al., EMNLP 2024), sentence-level, official chunking |
MIT |
| HHEM-2.1-Open | vectara/hallucination_evaluation_model |
Apache-2.0 |
| LLM judge | openai/gpt-oss-120b via Groq, asked for P(supported) |
API |
Limitations
- 40 test examples per domain: per-domain AUROCs have wide uncertainty.
- Very long CovidQA contexts are excluded (API limit of the LLM judge), so CovidQA results cover shorter documents only.
- HaluBench sources are public; detectors or the LLM judge may have seen some of them during training (possible contamination), which would inflate their scores on those sources.
About
Developed by Mohammed Achraf Larbi. The research framing (validate detectors before trusting them) was inspired by LLM-hallucination-Research by Dharambir Agrawal; all code here is original. Results generated 2026-10-03 12:34.