AI Paper Atlas

Research map / Vision-language and multimodal / CLIP and zero-shot

Locate then Correct: Debiasing Attention Heads in CLIP

Wei Jie Yeo, Rui Mao, Moloud Abdar, Erik Cambria, Ranjan Satapathy

ICML 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Fine-grained improvements to CLIP

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
add knowledge injection on the identified target states (ablation)LTC (MA), ViT-B/16 · Waterbirdsworst-group accuracy: 62.5 -> 73.3Improved
add mean-ablation of spurious states (ablation)LTC (KI), ViT-B/16 · Waterbirdsworst-group accuracy: 67.3 -> 73.3Improved
locate states with the contrastive method instead of random states (ablation)LTC (R), ViT-B/16 · Waterbirdsworst-group accuracy: 36.9 -> 73.3Improved
randomize the states ablated and injected (ablation)LTC, ViT-L/14 · Waterbirdsworst-group accuracy: 74.6 -> 15.2Hurt
apply knowledge injection to specific heads instead of the whole image representation (ablation)Roboshot (KI), ViT-B/16 · Waterbirdsworst-group accuracy: 45.6 -> 67.3Improved
mean-ablation of spurious states over zero-shot (ablation)ZS, ViT-B/16 · Waterbirdsworst-group accuracy: 49.7 -> 62.5Improved
tune the selected states on a validation set (LTC*)LTC, ViT-B/16 · Waterbirdsworst-group accuracy: 61.8 -> 73.3Improved
tune the selected states on a validation set (LTC*)LTC, ViT-H/14 · Waterbirdsworst-group accuracy: 73.7 -> 77.4Improved
apply LTC on top of a JTT-trained classifierJTT, ViT-B/16 · Waterbirdsworst-group accuracy: 72.3 -> 86.4Improved
apply LTC on top of a JTT-trained classifierJTT, ViT-H/14 · Waterbirdsworst-group accuracy: 85.7 -> 89.1Improved
replace orthogonal projection on the whole representation by LTCOrtho-Cali, ViT-H/14 · Waterbirdsworst-group accuracy: 18.1 -> 73.7Improved
apply LTC to multi-class counter-animal dataZS, ViT-B/16 · CounterAnimal hardaccuracy: 54.8 -> 55.2Improved

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