AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

Visual Interestingness Decoded: How GPT-4o Mirrors Human Interests

Fitim Abdullahu, Helmut Grabner

ICCV 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Human perception modeling · 0 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
restrict to images where humans reach consensusGPT-4o pairwise labels · Pairwise interestingnessagreement (%): 66.2% -> 73.8%Improved
evaluate on dissent set instead of full setGPT-4o pairwise labels · Pairwise interestingnessagreement (%): 66.2% -> 56.5%Hurt
swap image order and keep only order-consistent GPT-4o answersGPT-4o pairwise annotation · 2,500 Flickr image pairsfraction consistent: 36% always picked second image; 64% (1,599 of 2,500) consistentHurt
add demographic system prompt (gender, continent, age)GPT-4o · 116 image pairspreferenceNo clear effect
ask relative pairwise question instead of single-image questionGPT-4o · Flickr imageslabel informativenessImproved
use GPT-4o labels instead of aesthetic model labelspairwise labelling · Human labels agreementagreement (%): 68.3 -> 73.8Improved
use GPT-4o labels instead of CI model labelspairwise labelling · Human labels agreementagreement (%): 69.6 -> 73.8Improved
use GPT-4o labels instead of memorability labelspairwise labelling · Human labels agreementagreement (%): 35.5 -> 73.8Improved
use GPT-4o labels instead of CuPL zero-shot CLIPpairwise labelling · Human labels agreementagreement (%): 60.3 -> 73.8Improved
use GPT-4o labels instead of favorites countpairwise labelling · Human labels agreementagreement (%): 66.4 -> 73.8Improved
train ranker on GPT-4o labels instead of human labels, tested on human labelsCLIP learning-to-rank · Human test labelsaccuracy (%): 77.5 -> 73.4Mixed
use DINOv2 embeddings instead of CLIPlearning-to-rank · Human and GPT-4o labelsaccuracy (%): 67.1% for Humans and 68.3% for GPT-4oHurt

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