Implicit feedback from mouse and gaze used to align LLMs without explicit ratings
arXiv·medium signal
Chang and colleagues show that implicit user signals — mouse movement and eye/gaze behavior — leak user preferences and can align LLMs without collecting explicit human feedback or training a separate reward model. This suggests agent interfaces could learn from natural interaction telemetry rather than thumbs-up/down prompts. It carries an obvious privacy flip side: the same signals that improve alignment also reveal preferences users never explicitly shared.