Eye-Tracking Sharpens AI Reward Models

🔥 What happened A team led by Chaeho Lee trained an RLHF reward model using a spatially graded attention distribution from eye-tracking data, not just isolated fixations. The model beat the baseline on preference prediction. NeurIPS 2026 accepted the paper. 💡 Why it matters Prior gaze models treated each fixation as a separate point; this one weights surrounding text the way humans actually read. That could cut reliance on expensive preference labels. The accuracy bump shows cognitive psychology can directly improve alignment training. ⚡ Our take Finally a paper that models the human more precisely instead of just scaling compute. If your reward model only uses click data, you're leaving accuracy on the table.