arXiv · 2611.18670
Evidence-Calibrated Selective Decoding for Reliable Visual Question Answering
Abstract
Visual question answering systems can produce confident responses that are weakly supported by their input images. Contrastive decoding can reduce this dependence on language priors, but a fixed correction may also overturn a correct answer, while selective answering can conceal errors by reducing coverage. We develop evidence-calibrated selective decoding, a framework that separates three decisions: whether local visual evidence supports a candidate, how strongly to correct the original distribution, and whether the resulting answer should be released under a verification budget. Evidence Verification across Crops (EVC) measures visual support and disagreement; Adaptive Contrast Correction (ACC) applies a bounded modification within the backbone's plausible candidate set; Budgeted Selective Answering (BSA) calibrates an answer-admission threshold on held-out images. We provide a complete mathematical specification, bounded-perturbation analysis, and an evaluation protocol that distinguishes forced answering from selective risk. Our quantitative analysis re-examines published baselines on POPE instantiated over MSCOCO, A-OKVQA, and GQA, together with four MME subtasks. These public measurements reveal variation across backbones, tasks, and correction settings that motivates adaptive control. The proposed framework has not yet undergone model-level evaluation; its effectiveness remains a falsifiable hypothesis rather than an established empirical result.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Shiyu Gu, Luis Miguel Santana, Rafael Enrique Cruz. 2026-10-26. Evidence-Calibrated Selective Decoding for Reliable Visual Question Answering. https://searcharxiv.com/abs/2611.18670
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.