Cybersickness is the oldest unsolved problem in VR and the hardest to instrument, because the standard measurement is a questionnaire — and stopping someone to ask how nauseous they feel is both intrusive and a poor fit for anything you would want to do automatically.
Gaze Under Strain goes looking for the signal in the eyes.
It was posted to arXiv on 23 October 2026 (arXiv:2610.32019) by Nevzat Umut Demirseren, Mehedi Hasan Raju, Oleg Komogortsev and Isayas Berhe Adhanom. The Comments field reads “Accepted for publication at ACM VRST 2026. 10 pages, 4 figures.”
We checked the venue. VRST 2026 — the 32nd ACM Symposium on Virtual Reality Software and Technology — runs 16–18 November 2026 at Tohoku University in Sendai, Japan. So this is accepted work whose venue is three weeks away: timely, and with no DOI yet, so the arXiv record is the thing to cite.
Why the eyes are the right place to look
The paper’s mechanism argument is the reason this is not just a correlation hunt:
Arising from mismatches between visual and vestibular signals, it places sustained demands on gaze stabilization and visual processing.
Cybersickness is, in the dominant account, a conflict between what you see and what your inner ear reports. Gaze stabilisation is precisely the system that reconciles those two channels — the vestibulo-ocular reflex exists to hold your gaze still while your head moves. If the conflict is real, the machinery under load is oculomotor, which makes eye movement a plausible place for the strain to surface rather than an arbitrary biosignal that happens to correlate.
And the second-order point, which is more interesting:
Because gaze is the mechanism through which users actively sample the virtual environment, cybersickness-induced disruptions may alter both what users perceive and how discomfort itself emerges.
That is a feedback loop. Discomfort disturbs gaze; disturbed gaze changes what you look at; what you look at affects how sick you get. If that holds, cybersickness is not a passive symptom but partly self-reinforcing.
The study
- 19 participants
- a realistic virtual environment, navigated
- a 20-minute session
- discomfort rated repeatedly throughout, rather than once at the end
- linear mixed-effects models examining gaze constructs as functions of discomfort, exposure duration, and their interaction
Three families of gaze construct: fixation stability, visual exploration, and gaze allocation.
The repeated-ratings design is what makes the analysis possible. A single post-session score tells you who got sick; ratings through the session let you model discomfort as a trajectory and ask what the eyes were doing at each point. Mixed-effects models are the right tool, because each participant contributes many observations and people differ enormously in baseline susceptibility — a model that does not account for per-participant variation would attribute that to the effect.
What they found
Fixation stability degraded progressively with increasing discomfort, and the effect strengthened over time.
That interaction term is the headline. It is not merely that uncomfortable people hold fixation less well; it is that the relationship between discomfort and instability tightens the longer the exposure runs. Early in a session, discomfort and gaze are loosely coupled. Later, they track each other more closely.
Visual exploration showed no marginal discomfort effect but did show a significant effect involving discomfort — the arXiv abstract truncates mid-clause here, so we are not going to complete the sentence. The clean, quotable result is the fixation-stability one.
What it is and is not
19 participants. Small, though mixed-effects models on repeated measures extract considerably more from 19 people than a between-subjects design would. It is a within-subject result and should be read as one.
It is correlational. Discomfort was measured, not induced at assigned levels. Fixation instability and discomfort move together; the paper’s mechanism argument makes a causal story plausible, and the study does not establish direction. The feedback-loop framing is a hypothesis the design cannot test.
One environment, one session, no headset comparison. Which matters more than it sounds — a companion paper in the same arXiv batch (2610.32915) is specifically about how eye-tracking error varies across three VR headsets, which is a reminder that oculomotor measures are partly a property of the hardware measuring them.
No prediction claim. This is a behavioural analysis of how gaze evolves alongside discomfort. It is not a validated detector, and it does not report whether you could classify an individual’s discomfort from their gaze at a useful accuracy. That is the obvious next paper and this is not it.
Why it matters for anyone shipping XR
If fixation stability tracks discomfort, then a headset with eye tracking — which is now most of the premium ones — is already carrying the sensor you would need to notice a user becoming unwell without asking them.
That opens a design space worth thinking about carefully. An experience that detects degrading fixation stability could reduce optical flow, widen a comfort vignette, slow locomotion, or simply suggest a break — intervening before the user consciously registers feeling bad, which is the point at which intervention is cheap. Every VR comfort setting today is configured in advance by someone guessing; this is the input a system would need to adapt.
Two cautions on that enthusiasm. Continuous oculomotor monitoring is continuous biometric monitoring, and gaze data is unusually revealing — it carries attention, interest and, in the literature, correlates with clinical conditions nobody signed up to disclose to a game. And a detector built on a 19-person within-subject correlation is not a detector yet.
But as an argument that the signal exists and is tractable, this is a good piece of work, and the interaction with exposure time is the finding that would change how a system should behave: whatever you build should get more confident in the signal the longer someone has been in the headset.
Sources
- Gaze Under Strain: How Oculomotor Behavior Tracks Cybersickness Progression in VR — arXiv:2610.32019
- arXiv PDF
- ACM VRST 2026
- What Mean Error Misses: Device-Clock Diagnostics for Evaluating Eye Tracking across Three VR Headsets — arXiv:2610.32915