AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Trust but Verify: Programmatic VLM Evaluation in the Wild

Viraj Prabhu, Senthil Purushwalkam, An Yan, Caiming Xiong, Ran Xu

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 6 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
larger InternVL2 model size 2B to 26BInternVL2 · PROVEhscore: 73.96 -> 74.63Improved
larger InternVL2 model size 2B to 26BInternVL2 · PROVEtscore: 79.51 -> 79.23Hurt
scale LLaVA-1.5 from 7B to 13BLLaVA-1.5 · PROVEtscore: 82.58 -> 82.40No clear effect
scale LLaVA-1.5 from 7B to 13BLLaVA-1.5 · PROVEhscore: 72.67 -> 72.46No clear effect
newer LLaVA-Next-7B instead of LLaVA-1.5-7BLLaVA · PROVEhscore: 72.67 -> 74.28Improved
newer LLaVA-Next-7B instead of LLaVA-1.5-7BLLaVA · PROVEtscore: 82.58 -> 80.03Hurt
give the model the ground-truth caption (oracle) versus GPT-4oLLaMA-3.1-8B oracle · PROVEaverage: 78.72 -> 84.22Improved
GPT-4o versus Claude-3.5-Sonnetclosed-source VLMs · PROVEtscore: 77.31 -> 80.92Improved
InternVL2-26B versus simpler LLaVA-1.5-7B on truthfulnessVLMs · PROVEtscore: 82.58 -> 79.23Hurt
correlation between helpfulness and truthfulness across modelsVLMs · PROVEPearson correlation: 0.03No clear effect
compare automatic metrics to human ratingsPROVE metrics · PROVE human studyPearson correlation: 0.81 vs 0.45Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy