AI Paper Atlas

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

SugarCrepe: Fixing Hackable Benchmarks for Vision-Language Compositionality

Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, Ranjay Krishna

NeurIPS 2023 · Vision-language and multimodal › Human and embodied modeling · approach: Human-like visual cognition in VLMs · 25 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
Replace rule-based negative generation with ChatGPT generationCompositionality benchmark construction · ARO+CREPE vs SugarCrepe hard-negative textcommonsense and grammar quality: 86%Improved
Adversarially refine the generated benchmark to balance text-model score gapsSugarCrepe · SugarCrepetext-only attack accuracy: 0.5Improved
Train with Replace hard negativesViT-B/32 NegCLIP finetuning · ARO+CREPE Replaceaccuracy: 77.06 -> 94.51Improved
Train with Replace hard negativesViT-B/32 NegCLIP finetuning · SugarCrepe Replaceaccuracy: 84.76 -> 88.27Improved
Train with Swap hard negativesViT-B/32 NegCLIP finetuning · ARO+CREPE Swapaccuracy: 68.81 -> 94.48Improved
Train with Swap hard negativesViT-B/32 NegCLIP finetuning · SugarCrepe Swapaccuracy: 70.83 -> 76.21Improved
Evaluate hard-negative training on less biased SugarCrepe rather than ARO+CREPENegCLIP finetuning · Replace hard negativesaccuracy gain over CLIP finetuning: 17.45; 3.51Mixed
Scale pretrained CLIP models and training dataPretrained CLIP · SugarCrepe Swap-Obj and Swap-AttaccuracyNo clear effect
Use object changes rather than attribute or relation changesPretrained CLIP models · SugarCrepe Replaceaccuracy: 15%; 29%Improved
Use object additions rather than attribute additionsPretrained CLIP models · SugarCrepe Addaccuracy: 10%Improved

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