Cognitive Gaze Augmented Reasoning: Teaching V-Thinker to “See Like a Human”

TL;DR: By modeling and incorporating human attention scanpaths (how people look at images), let’s see if V-Thinker can reason more naturally and efficiently—especially in tasks requiring fine spatial analysis. The initial experiment overlays radiologist or crowdsourced eye-tracking data as soft-attention supervision during training.

Research Question: Does augmenting interactive image reasoning models with real human gaze data (scanpaths, fixations) improve their region selection and sequential reasoning, bridging the gap between human and machine visual cognition?

Hypothesis: Incorporating human gaze dynamics—either as soft attention priors or curriculum cues—will guide V-Thinker-like models toward more efficient and human-aligned visual reasoning, boosting both interpretability and accuracy on challenging visual tasks.

Experiment Plan: - Source or create gaze datasets (e.g., MedGaze, crowdsourced scanpaths) for complex images.

• Modify V-Thinker’s Data Evolution curriculum to include gaze-based region prioritization or sequence constraints.
• Train and evaluate on reasoning tasks—especially ones requiring spatial search or anomaly detection.
• Compare to standard V-Thinker on reasoning efficiency, accuracy, and alignment with human explanations.

References: ['Qiao, R., Tan, Q., Yang, M., Dong, G., Yang, P., Lang, S., Wan, E., Wang, X., Xu, Y., Yang, L., Sun, C., Li, C., & Zhang, H. (2025). V-Thinker: Interactive Thinking with Images.', 'Awasthi, A., Le, N., Deng, Z., Agrawal, R., Wu, C. C., & Nguyen, H. (2024). Multimodal Learning and Cognitive Processes in Radiology: MedGaze for Chest X-ray Scanpath Prediction. arXiv.org.', 'Chaudhari, S., Akula, T., Kim, Y., & Blake, T. (2025). Multimodal LLM Augmented Reasoning for Interpretable Visual Perception Analysis. arXiv.org.']