Are Hallucination Probes Evaluated Correctly? A Critical Review and a Pilot Study under 4-bit Quantization
Year: 2026
Type: preprint
Abstract: Linear probes trained on a language model's hidden states are among the cheapest hallucination detectors, and many papers report AUROC above 0.9. A growing body of 2025–2026 work suggests these numbers often measure something other than the ability to tell a correct answer from a wrong one. We review how hallucination probes are evaluated and organize the known failure modes into five groups: noisy correctness labels, pooled metrics that mix question difficulty with answer correctness, question-side shortcuts, text artifacts that simple baselines exploit, and weak generalization. We then examine a deployment setting the literature has barely addressed: probes trained at full precision but used on quantized models. Existing quantization studies report only pooled metrics, so they cannot show whether the answer-level signal survives compression. We propose a 12-item reporting checklist and apply it in a pilot study on Qwen2.5-1.5B-Instruct with 1,000 TriviaQA questions at FP16, INT8 and NF4. A question-only probe reaches 0.73 pooled AUROC without seeing any answer, against 0.85 for the answer probe, so most of the pooled signal reflects question difficulty. The within-question AUROC of 0.76 confirms a genuine answer-level signal, and it did not measurably change under NF4 (difference −0.002, 95% CI [−0.08, 0.09]). The pilot is small, and its confidence intervals rule out only large degradations; we report it as a worked example of the checklist rather than a definitive result.
Author: Sharma, Mritunjay
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7228099144
