Varun Kotte · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3832562
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
The prevailing assumption in the multimedia community is that scaling will eventually solve the hallucination problem in Vision-Language Models (VLMs): larger models will generate more accurate outputs and better verify them. We challenge this assumption with both empirical evidence and a theoretical argument. Through systematic experiments across three VLM scales (3B, 7B, 72B parameters) on document understanding tasks, we demonstrate that while generation accuracy improves substantially with scale (+26pp from 3B to 72B), verification quality actually degrades: error detection rate drops from 43% to 8% at the 7B scale, with the smallest model proving the best error detector. We observe that model confidence is anti-predictive of verification quality: false acceptances tend to carry higher confidence than correct rejections, undermining confidence-based filtering strategies. We explain these findings through the generation-verification asymmetry thesis: generation compresses input evidence into outputs, while verification requires decompressing outputs back to evidence, a capability that current transformer architectures lack and that scaling does not provide. This asymmetry is amplified in VLMs, where visual evidence undergoes additional lossy compression through vision encoders. As a proof-of-concept, we show that chain-of-thought reasoning provides zero FAR improvement, while evidence-citing verification (forcing models to locate and quote input evidence before judging) reduces FAR by up to 19pp at the 72B scale, albeit at the cost of higher false rejection, suggesting that verification-native architectures, not scale alone, are the path forward. We conclude by proposing five open problems that frame verification as a first-class research challenge for the multimedia community.
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
No comments yet — start the discussion below.