AI Paper Atlas

Research map / Large language models / Alignment and preferences

Unified Generation and Self-Verification for Vision-Language Models via Advantage Decoupled Preference Optimization

Xinyu Qiu, Heng Jia, Zhengwen Zeng, Shuheng Shen, Changhua Meng, Yi Yang, Linchao Zhu

CVPR 2026 · Large language models › Alignment and preferences · approach: Robust and calibrated DPO · 7 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 binary verification reward with preference verification reward (ablation)ADPO · AndroidControlAUC: +19.4%Improved
replace binary verification reward with preference verification reward (ablation)ADPO · MathVistabest@8 accuracy: +1.6%Improved
replace binary verification reward with preference verification reward (ablation)ADPO · MathVistaAP: +1.3%Improved
use decoupled advantage instead of entangled advantage (ablation)ADPO · AndroidControlAUC: +34.1%Improved
use decoupled advantage instead of entangled advantage (ablation)ADPO · GUI agent tasksbest@8 performance: +2.8%Improved
naively sum answer and verification rewards (entangled advantage) (ablation)ADPO · traininggeneration qualityHurt
use binary verification reward under class imbalance (ablation)ADPO · trainingverification score discriminationHurt
set preference margin gamma to 0.1 instead of 0.025 (ablation)ADPO · ReasonSeg (overall)ACC: 71.1 -> 73.5Improved
set preference margin gamma >= 0.2 (ablation)ADPO · ReasonSeg (overall)ACC: 73.5 -> 72.3Hurt
use unified ADPO policy as verifier instead of GRPO-as-judgeGRPO generator · MathVista best-of-8accuracy (%) / latency: 60.8 -> 65.0 acc; 5.6 -> 2.6 sImproved
ADPO training vs GRPO on pass@1ADPO · MMMU (OOD)accuracy (%): 48.7 -> 47.7No clear effect
ADPO training vs GRPO on pass@1ADPO · ReasonSeg (overall)cIoU: 59.5 -> 59.1No 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