Vision-language models (VLMs) that receive an image and a text prompt can produce seemingly sophisticated responses, yet they often hallucinate objects, posing risks for safety‑critical use cases. Some detection approaches tap into the model’s internal signals during generation, sometimes reusing the very signals that caused the hallucination. We conduct a thorough analysis of the robustness of these detection techniques.
First, using the token position of a generated word as a feature yields performance comparable to more elaborate methods, indicating that position alone is a strong heuristic. Further experiments reveal that several methods rely on this positional cue as a shortcut rather than on genuine visual grounding.
Second, we examine sensitivity to language versus vision modalities. Findings suggest that strong detection performance is largely driven by spotting language‑based hallucinations; methods that depend on token position tend to be less visually grounded and lose robustness when hallucinations appear early in the text.
Finally, we show that different detectors carry complementary signals—adding weaker methods to an ensemble of stronger ones still improves overall results. In sum, internal VLM signals can aid hallucination detection, but existing approaches exhibit notable limitations.
Review: While internal signals show promise, future work must reduce reliance on token order and strengthen visual grounding to achieve reliable detection in real‑world safety scenarios.