New GazeQwen model enables AI to understand videos by tracking where users are looking, achieving 63.9% accuracy on eye-gaze tasks.

GazeQwen adds gaze awareness to Qwen2.5-VL-7B through a lightweight gaze resampler with only 1-5 million trainable parameters

Uses eye-fixation data and V-JEPA 2.1 video features injected into LLM decoder layers via hidden-state modulation for efficient gaze integration

Achieves 16.1 percentage point improvement over the base Qwen model with gaze prompts and outperforms GPT-4o on the StreamGaze benchmark

Optional second training stage adds low-rank adapters (LoRA) for tighter LLM integration and improved performance