Does gaze reveal whether people have achieved common ground?

The paper asks "whether gaze provides evidence about grounding" in collaborative tasks where participants hold different private information, and reports that it does, in the same direction across two corpora. In HCRC MapTask, aligned reference interpretations, and in MUNDEX, UND (understood) judgments, are associated with more task-directed gaze and less partner-directed gaze, lower gaze entropy and fewer gaze transitions. The associations are clearest for the participant leading the task: giver-produced references in MapTask and explainer judgments in MUNDEX, the latter also co-varying with the explainee's gaze. Follower-produced references show "near-zero effects."

The mechanism the authors give is the standard grounding account. When information is asymmetric, mutual understanding "cannot be assumed from shared context alone" and must be built and tracked through interaction, citing Clark and Brennan. Gaze is offered as an observable cue to that process, since participants look at task materials, at each other, or away while instructing, checking understanding and coordinating perspectives. The direction of the result echoes earlier map-task observations that partner-directed gaze rises around communicative difficulty. One within-speaker result stands out: in MapTask reference chains, the speaker's gaze entropy is lower at the mention where a previously nonaligned referent becomes aligned, which puts the signal at the moment of repair rather than only across whole conversations.

Against the library, this gives Why do speakers need to actively calibrate shared reference? a nonverbal trace. That note holds that shared surface words guarantee nothing about shared referents; here the evidence that calibration succeeded shows up partly outside the words. It also complicates Why don't conversational AI systems mirror their users' word choices?, which treats alignment as visible in lexical choice; gaze is a second, non-lexical channel of the same alignment work. The predictive gain is modest, which sits well beside Can conversation structure predict dialogue success better than content?: both find signal in how an interaction unfolds rather than what is said, but the gaze features improve only "modestly over controls under grouped cross-validation," and the two studies predict different targets, so the comparison is loose.

The excerpt is silent on sample sizes, effect sizes, causal direction and whether any of this helps a model. It flags its own limits. The two labels "measure different constructs," referential alignment in MapTask against pooled explainee self-reports and explainer judgments in MUNDEX, so the convergence is across related but distinct measures. Which features predict best also differs, with temporal dynamics in MapTask and raw proportions in MUNDEX. The shared partner/task/away categories name gaze targets, not functions: a partner glance may check a landmark reference, and averted gaze has been linked to topic changes. What follows at this strength is narrow: gaze is weak but consistent evidence of grounding in humans, concentrated in whoever leads. A text-only interface has no such channel, so a system there gets grounding evidence only from the words.