In collaborative tasks where each participant holds different information — maps with different landmarks, or a board game only one side can see — mutual understanding must be built through interaction. A new paper asks whether gaze offers observable evidence of that process across two such tasks.
The study compares HCRC MapTask, where a reference is marked as aligned only when speaker and addressee interpretations match, and MUNDEX, a corpus of German game explanations with retrospective understanding judgments from both sides. Both corpora annotate gaze as discrete behavioral categories rather than eye-tracking coordinates, but use different category sets (up/down/off versus partner/table/away). The authors map them into a shared partner/task/away vocabulary and compute gaze features around each grounding-labeled dialogue unit.
Across both corpora, the associations point in the same direction: aligned references and "understood" judgments come with more task-directed gaze, less partner-directed gaze, lower gaze entropy, and fewer gaze transitions. The pattern is clearest for the participant leading the task — givers in MapTask and explainers in MUNDEX — and explainer judgments also co-vary with the explainee's gaze. In same-speaker MapTask reference chains, speaker gaze entropy drops at the mention where a previously non-aligned referent becomes aligned.
Prediction gains remain modest: the best gaze feature groups improve only slightly over role and condition controls under grouped cross-validation, with temporal features performing best in MapTask and raw proportions in MUNDEX. The authors also note that several effects weaken when recurring participants, rather than dialogues, are used as the unit of inference.
Because the effects are small, the paper treats gaze as one contributing cue to grounding rather than a standalone signal, to be interpreted alongside task and dialogue context. The authors suggest that since the representation requires only discrete gaze labels, it should port to other corpora, and invite discussion on whether shared category names pick out the same interactional function across quite different tasks.