AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Towards Robustness against Typographic Attack with Training-free Concept Localization

Bohan Liu, Wenqian Ye, Guangzhi Xiong, Zhenghao He, Sanchit Sinha, Aidong Zhang

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 1 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
add lexical-circuit intervention (attention reweighting) to CLIPViT-H/14 · RTA-100object classification accuracy: 53.4 -> 76.2Improved
add lexical-circuit intervention to CLIPViT-H/14 · RTA-100text confusion rate: 42.0 -> 14.4Improved
add lexical-circuit intervention to CLIPViT-B/16 · Disentanglingobject classification accuracy: 52.2 -> 88.3Improved
add lexical-circuit intervention to CLIPViT-B/16 · IN-100-Texttext confusion rate: 33.0 -> 7.1Improved
use nTAS circuit mining instead of DyslexifyCLIP ViTs · RTA-100 (avg)zero-shot accuracy: 67.0 -> 71.7Improved
use nTAS circuit mining instead of DyslexifyCLIP ViTs · Disentangling (avg)zero-shot accuracy: 67.6 -> 78.7Improved
use nTAS circuit mining instead of Defense-PrefixCLIP ViTs · IN-100-Text (avg)zero-shot accuracy: 70.1 -> 77.0Improved
use nTAS intervention instead of Defense-PrefixCLIP ViTs · ImageNet-100 (avg)clean accuracy: 81.4 -> 81.0No clear effect
add zero ablation of lexical circuitGemma3-12B · RIO-Bench obj-attackVQA accuracy: 48.24 -> 49.78Improved
add zero ablation of lexical circuitQwen3-VL-8B · RIO-Bench obj-attackVQA accuracy: 70.11 -> 71.69Improved
add zero ablation of lexical circuitInternVL3.5-14B · RIO-Bench obj-attackVQA accuracy: 53.07 -> 53.08No clear effect
add zero ablation of lexical circuitvision encoders of LVLMs · RIO-Bench cleanVQA accuracy: -0.3% to +0.6%No clear effect

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