AI Paper Atlas

Research map / Adversarial robustness and security / Attacks on text and graphs

AdPO: Enhancing the Adversarial Robustness of Large Vision-Language Models with Preference Optimization

Chaohu Liu, Gui Tianyi, Yu Liu, Linli Xu

ICLR 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Vision-language model robustness

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
use AdPO-trained CLIP encoder instead of FARE or TeCoA encoders in LVLMsLLaVA 1.5-7B · COCO captioning l_inf 2/255CIDEr: FARE2 53.6 -> AdPO2 62.7 adversarial; clean 109.9 -> 114.3Improved
train robust encoder on small TinyLLaVA (OpenELM-450M) and transfer to OpenFlamingo-9B and LLaVAAdPO · OpenFlamingo-9B COCOCIDEr: OF-9B COCO clean 79.6, 2/255 38.2 vs FARE2 34.2Improved
increase training perturbation radius from 2/255 to 4/255 (ablation)AdPO · LLaVA 1.5-7B COCOCIDEr: clean 114.3 -> 109.2; l_inf 4/255 40.7 -> 47.8Mixed
swap DPO for SimPO (no reference model) (ablation)AdPO · VQAv2VQA accuracy clean / adversarial: SimPO 68.1 / 30.7; IPO 69.2 / 32.6Hurt
swap DPO for IPO, KTO or StepDPO (ablation)AdPO · VQAv2VQA accuracy: clean 69.2-70.1; adversarial 32.0-32.6No clear effect
evaluate against C&W or sC&W attacks instead of APGD (ablation)AdPO · VQAv2adversarial accuracy: APGD 32.4; C&W 33.1; sC&W 32.8No clear effect
increase lambda weight on adversarial image optimization (ablation)AdPO · VQAv2clean vs adversarial accuracyMixed
use AdPO encoder against targeted attacksLLaVA 1.5-7B · Targeted attack, eps=4/255mean ASR: 100% (CLIP) -> 0% (AdPO2)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