Research map / Vision-language and multimodal / Multimodal LLMs

A3: Towards Advertising Aesthetic Assessment

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces A3, a framework with a three-stage advertising aesthetic paradigm (A3-Law), a 30K-image/120K-pair CoT dataset, an MLLM (A3-Align, Qwen3-VL-8B trained with SFT then GRPO with multi-signal rewards), and a benchmark against 20+ MLLMs.

Problem

Advertising image aesthetic evaluation relies on subjective human judgment and one-step holistic MLLM scoring, which lack scalability, standardized criteria, interpretability, and stepwise diagnostic feedback.

Approach

A theory-driven hierarchical paradigm (Perceptual Attention, Formal Interest, Desire Impact) inspired by AIDA is operationalized into rule-based annotations, CoT rationales, and rule-specific rewards (IoU, Gaussian continuous score, tool use) for GRPO training of an MLLM.

Key results

• A3-Align reaches 0.926 accuracy and 0.701 mAP@0.5 on promotional iconography, more than double the best closed-source mAP (0.283).
• A3-Align achieves 0.880 SRCC on Aesthetic Attribute, above the best baseline Claude Opus 4.1 (0.772).
• In quality ad selection, A3-Align raises Satisfaction by +1.05 (35.5%) and Action Intent by +1.20 (40.1%) at Desire Impact stage; in critique, it averages 4.65 across expert axes, +0.37 over next best.

Surprising findings

• Enabling thinking mode is unstable or harmful for open-source models, while closed-source models consistently benefit.
• Both open- and closed-source MLLMs struggle with promotional icon localization, with best closed-source mAP of only 0.283.

Limitations

• Failure on dense layouts causing icon hallucinations or merging.
• Underestimates minimalist premium designs by over-associating simplicity with weaker advertising effectiveness.
• Cultural calibration of the paradigm left for future work.

Experimental setup

Datasets: A3-Dataset AdImageNet

Benchmarks: A3-Bench AdImageNet

Metrics: Accuracy mAP@0.5 SRCC PLCC Satisfaction and Action Intent gains expert ratings of problem identification, CoT depth, clarity

Baselines: Qwen3-VL-8B-Instruct Gemma-3-27B-it Qwen3-VL-32B-Instruct Qwen2.5-VL-72B-Instruct GLM-4.5V Llama-4-Scout Llama-4-Maverick GPT-5 Gemini 2.5 pro Claude Opus 4.1

Builds on: GRPO AIDA model Qwen3-VL Hasler colorfulness

Closest papers by meaning

• AI-generated Image Quality Assessment in Visual Communication (AAAI 2025 · similarity 0.86)
• GeoReward: Mitigating Contextual Variable Overestimation in Vision-Language Models for Cross-Market Preference Prediction (ICML 2026 · similarity 0.84)
• Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.83)
• ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding (CVPR 2026 · similarity 0.83)
• Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.83)
• UniAR: A Unified model for predicting human Attention and Responses on visual content (NeurIPS 2024 · similarity 0.83)