LLM Hallucination Detection Hits F1 0.915 — In-Domain Only

A preprint posted yesterday (arXiv:2609.11878) reports a three-signal hallucination detection pipeline scoring F1 0.915 and AUROC 0.977 on HaluEval. Read the second half of the abstract before you integrate it: on the biomedical SciFact benchmark, the same general-domain pipeline scores F1 0.52. The headline number and the cautionary number come from the same model.

The problem: catching unfaithful claims after generation

Hallucination detection has been stuck between two bad options: ask the generating model to self-assess (it's confidently wrong at exactly the moments that matter), or run expensive uncertainty machinery over every token. What practitioners actually need is a cheap, external judge that flags a whole response before it ships. This paper targets that slot.

The method: three cheap signals stacked

The pipeline combines a fine-tuned DeBERTa-v3 classifier, Monte Carlo Dropout uncertainty at inference, and temperature-scaled calibration. Nothing exotic — the contribution is the composition. The ablation that matters: removing the knowledge context from summarization inputs drops F1 by 24%, which is evidence the model is doing genuine entailment against provided context rather than pattern-matching on fluency or surface cues. Also worth stealing: 25% of the training data recovers 77% of full-data performance. You can bootstrap a usable domain detector from a few thousand labeled samples.

mindmap
  root((Hallucination
Detector))
    DeBERTa-v3 classifier
      F1 0.915 HaluEval
      QA 0.97 / Summ 0.96 / Dialogue 0.82
    MC Dropout uncertainty
      accuracy 88.7 → 93.2%
    Calibration
      temperature-scaled
    Domain collapse
      SciFact F1 0.52 general
      SciFact F1 0.63 PubMedBERT
            

The results, and what they actually mean

Three numbers deserve scrutiny beyond the abstract:

• 0.915 F1 is a benchmark number, not a deployment number. Per-task scores tell the real story: QA 0.97, summarization 0.96, but dialogue only 0.82. HaluEval's hallucinations are constructed perturbations of grounded text. Dialogue hallucinations — omissions, wrong entities, subtle drift — are harder, and that's the regime production chatbots live in.
• The DPO result is measured by the detector itself. Fine-tuning Qwen2.5-0.5B with DPO cut hallucination rate from 85.5% to 37.7% — a 55.9% relative reduction — as scored by the same detector that was trained on HaluEval. That's a circular measurement: if the detector has blind spots, the DPO training optimizes into those blind spots. The 37.7% figure should be treated as "detector-visible hallucinations dropped," not as ground truth.
• The cross-domain collapse is the real finding. General-domain training transfers at F1 0.52 to SciFact — barely above coin-flip territory for a binary judge. Domain-matched PubMedBERT fine-tuning recovers to 0.63. Read that carefully: even the best adaptation strategy lands 28 points below the in-domain headline. Nobody deploying this in medicine, law, or finance gets 0.915. They get 0.63 at best, after fine-tuning.

Limitations the authors downplayed

The paper doesn't state the false-positive rate at operating thresholds, which is the number that decides whether you can auto-block flagged responses or must route them to human review. A detector with AUROC 0.977 can still be unusable at the precision a moderation pipeline needs. Second, all results are response-level classification on benchmark data — no test on live model outputs where hallucination prevalence is much lower than HaluEval's roughly balanced classes, which will crush precision. Third, the MC Dropout gain (to 93.2% accuracy) comes at ~2x inference cost, which the paper doesn't price. And the detector targets the 0.5B–7B open-model class; whether the same signals transfer to frontier models that hallucinate differently is untested.

Why builders should care

The actionable takeaway isn't the architecture — it's the transfer result. If you're adding a hallucination guardrail to a domain-specific product, do not fine-tune on generic hallucination corpora and expect it to work. Budget for a few thousand domain-labeled examples (the 25%-of-data learning curve says that's enough for ~77% of ceiling), keep the detector's training distribution matched to your generation task, and — critically — if you use the detector to train the generator via DPO, validate the improvement with an independent judge or you're measuring your own blind spots.