# PRE-MAP: Personalized Reinforced Eye-tracking Multimodal LLM for High-Resolution Multi-Attribute Point Prediction

**Authors:** Hanbing Wu, Ping Jiang, Anyang Su, Chenxu Zhao, Tianyu Fu, Minghui Wu, Beiping Tan, Huiying Li  
**arXiv:** [2507.19213](https://arxiv.org/abs/2507.19213)  
**Date:** 2025-07-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2507.19213v1/

## Key points

- Individual preferences are crucial in determining how humans prioritize visual stimuli in the realm of visual selective attention.
- Existing models often overlook the impact of subjective cognitive diversity on fixation behavior.
- Conventional saliency prediction models have limitations in capturing personalized attention patterns due to reliance on low-resolution imagery and subsequent upscaling.
- A new approach called SPA-ADV has been introduced, involving a large-scale multimodal dataset and a novel eye-tracking saliency model known as PRE-MAP.
- The PRE-MAP model aims to characterize personalized visual disparities through Reinforcement learning-optimized Eye-tracking and predict format-correct and spatially accurate points guided by Multi-Attribute user profiles.
- C-GRPO has been introduced to enhance MLLMs' performance in producing precise prediction points while considering variability in eye movement points and Multi-Attribute profiles.
- Extensive experiments have demonstrated the effectiveness of these approaches in addressing challenges related to personalized gaze prediction within eye-tracking models.

## Summary

, , , , 

In the realm of visual selective attention, individual preferences play a crucial role in determining how humans prioritize visual stimuli. By bridging subjective cognitive mechanisms with objective visual elements, individuals are able to regulate their prioritization of dynamic visual scenes, thereby influencing semantic interpretation and hierarchical processing. However, existing models and datasets often overlook the impact of subjective cognitive diversity on fixation behavior. Conventional saliency prediction models typically rely on segmentation approaches with low-resolution imagery to generate saliency heatmaps that are subsequently upscaled to native resolutions, limiting their ability to capture personalized attention patterns. Moreover, Multimodal Language Models (MLLMs) face constraints such as hallucinations, making it challenging to adhere strictly to expected formats in tasks involving multiple point predictions. Achieving precise point positioning is also a significant challenge for these models. To address these limitations, a new approach called Subjective Personalized Attention for Advertisement Videos (SPA-ADV) has been introduced. This approach involves a large-scale multimodal dataset capturing gaze behaviors from over 4,500 participants across varying age groups and genders while analyzing 486 videos. Additionally, a novel eye-tracking saliency model known as PRE-MAP has been proposed to characterize personalized visual disparities through Reinforcement learning-optimized Eye-tracking. Built upon MLLMs and guided by Multi-Attribute user profiles to predict Points accurately, this model aims to ensure format-correct and spatially accurate prediction points are generated. To further enhance the performance of MLLMs in producing precise prediction points while considering the variability in eye movement points and Multi-Attribute profiles, Consistency Group Relative Policy Optimization (C-GRPO) has been introduced. Extensive experiments conducted on SPA-ADV and other benchmarks have demonstrated the effectiveness of this approach in addressing the aforementioned challenges. The code and dataset associated with this research are available at the provided URL. The study was authored by Hanbing Wu, Ping Jiang, Anyang Su, Chenxu Zhao, Tianyu Fu, Minghui Wu, Beiping Tan, and Huiying Li. This work represents a significant advancement in personalized gaze prediction within eye-tracking models.