Avatar eyes in social VR: balancing gaze communication and uncanny-valley effects

Introduction Eye gaze is a core channel of nonverbal communication, yet designing an embodied virtual agent (EVA) that conveys gaze clearly in virtual reality (VR) without inducing discomfort remains challenging. We investigated how explicitly depicting the current gaze target within the eye region affects (i) gaze decoding (social information transfer), (ii) avatar impression and perceptual comfort (uncanny valley related responses), and (iii) social presence during interaction. Methods We conducted a within-subject VR study ( N = 24) using a Varjo XR-3. Participants experienced three EVA eye designs: Pupil (minimally augmented eyes without explicit target depiction), Blend (an iris scaffold retained while the gaze target is depicted in a reflection-like manner), and Projection (iris/pupil removed and the current gaze target projected onto the eye region). Task 1 was an object-based gaze-identification task measuring accuracy and decision time, followed by a brief Likert questionnaire on perceived readability and eye impression. Task 2 was a joint-attention interaction in which the EVA contingently followed the participant’s gaze, followed by a brief Likert questionnaire on joint-attention experience, co-presence, perceived contingency, and eye impression. After all conditions, participants completed a comparative post-experiment questionnaire. Results Increasing gaze-target explicitness improved gaze decoding: Projection yielded the highest objective accuracy and was most frequently selected as the easiest condition for understanding the EVA’s gaze, while Blend and Projection improved perceived readability relative to Pupil . However, explicit projection incurred perceptual costs: Projection was rated as the least natural and was overwhelmingly selected as the most uncanny, whereas Pupil was rated as the most natural. Blend was most frequently preferred overall, indicating that users often favoured a compromise between readability and perceptual comfort. Social-presence-related ratings in Task 2 did not reliably differ across eye designs when gaze behaviour was held constant. Conclusion Making gaze targets explicit within the eye region can improve gaze communication in VR but risks discomfort from perceptual mismatch. Hybrid designs that embed target information within an eye-like appearance may offer a practical balance between communicative utility and perceptual comfort for social VR agents.