Answers are grounded in this paper and cite their sources.
Inspectable visual evidence does not necessarily indicate that an evidence-based acceptance rule is more reliable in prediction. The aim of this study is to measure that difference in binary object-presence visual question answering. It is accompanied with an independent object detector providing localized evidence for positive answers, and an open multimodal generator, SmolVLM-256M-Instruct. The number of positive candidates accepted depends on a development-selected threshold, while negative candidates abstain because lack of detection cannot establish absence. The exploratory pilot resulted in a new replication with 60 new COCO val2017 images and 360 new adversarial POPE questions, where the images were split into development/evaluation sets. 68.75% reference-label agreement on 240 evaluation questions was obtained using direct generation. The evidence gate accepted 39 answers at 16.25% coverage with 100.00% reference agreement, while positive-only confidence selection achieved 97.44% agreement at the same accepted count. The gate had a higher accepted-answer accuracy point estimate than positive-only confidence selection, but the exploratory difference interval included zero; superiority was not established. Also reported in this study are threshold sensitivity, positive-answer retention, image-cluster bootstrap intervals and checks on component training provenance. The results apply to just one backbone and a limited object-presence task, and do not establish semantic evidence sufficiency, causal visual dependence or general hallucination mitigation. The contribution is a reproducible diagnostic case study that demonstrates how answer-class composition and component training provenance should be controlled when assessing the evidence-constrained visual answering problem.
No takes yet. Share an insight, caveat, or question.
Ahmed et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: