Research map / Vision-language and multimodal / Multimodal LLMs

GeoReward: Mitigating Contextual Variable Overestimation in Vision-Language Models for Cross-Market Preference Prediction

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

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.

Identifies Contextual Variable Overestimation in VLMs and proposes GeoReward, a Qwen2-VL-7B-based reward model with market-aware retrieval, country-conditioned visual modulation, and selective sensitivity loss for cross-market ad preference prediction, which also guides DPO fine-tuning of a background-design generator.

Problem

VLMs overweight dominant visual-textual cues and underweight sparse but decision-critical context variables (e.g., country), causing market-invariant predictions in cross-market ad image preference prediction.

Approach

Combines three components: retrieval of region-aligned CTR-based preference evidence (MA-RAG) injected as a prior, FiLM-style scale and shift on visual tokens generated from the country embedding (CGVM), and a loss that penalizes errors more when the model under-attends to country, product, and image tokens (SSL); the trained reward model then drives DPO on a background design generator.

Key results

• Full GeoReward reaches 60.37% accuracy and 40.84% sensitivity on MACP, versus 55.60% and 36.73% for Qwen2-VL-7B with an FC head.
• On the transformed ImageGen-CoT-Reward-5K task, GeoReward reports 67.9% accuracy versus 62.7% for finetuned UnifiedReward.
• A reported online A/B test over one week and 6 million impressions shows average CTR gains of 1.1% for images from the RL-tuned design generator.

Surprising findings

• The plain finetuned Qwen2-VL-7B collapses to near-chance accuracy (44.61%) on MACP, consistent with the claimed CVE effect.
• Embedding-based and instruction-based RAG augmentation hurt performance relative to the FC-head baseline until the augmented embedding is rescaled.

Limitations

• Targets a single sparse variable (country); multiple simultaneous sparse variables are not evaluated.
• No comprehensive benchmark across diverse domains beyond the ad dataset and one transformed image-quality task.
• Most errors occur when candidate images are visually near-identical.

Experimental setup

Datasets: MACP ImageGen-CoT-Reward-5K

Benchmarks: MACP ImageGen-CoT-Reward-5K

Metrics: accuracy sensitivity

Baselines: Qwen2-VL-7B (finetuned) Qwen2-VL-7B (with FC Head) UnifiedReward

Builds on: Qwen2-VL DPO ControlNet

Closest papers by meaning

• Visual Persuasion: What Influences Decisions of Vision-Language Models? (ICML 2026 · similarity 0.87)
• Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.86)
• Personalized Preference Fine-tuning of Diffusion Models (CVPR 2025 · similarity 0.85)
• Importance Sampling for Multi-Negative Multimodal Direct Preference Optimization (ICLR 2026 · similarity 0.84)
• RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data (AAAI 2025 · similarity 0.84)
• Zooming from Context to Cue: Hierarchical Preference Optimization for Multi-Image MLLMs (NeurIPS 2025 · similarity 0.84)