Gaze as Evidence for Common Grounding: A Cross-Corpus Analysis of MapTask and MUNDEX
Nan Li, Albert Gatt, Massimo Poesio
Utrecht University · Queen Mary University of London
Published September 16, 2026 · on the feed September 17, 2026 · ▲ 30 · arXiv 2609.18011
The abstract
In collaborative tasks with asymmetric information, participants coordinate their understanding through interaction. We ask whether gaze provides evidence about grounding across two such tasks. Working from discrete behavioral annotations, we map HCRC MapTask (Anderson et al., 1991) and MUNDEX (Türk et al., 2023) into a shared partner/task/away vocabulary and compute gaze features around task-relevant dialogue units. In both corpora, aligned reference interpretations (MapTask) and UND (understood) judgments (MUNDEX) are associated with more task-directed gaze and with less partner-directed gaze, lower gaze entropy, and fewer gaze transitions. The associations are clearest for the participant leading the task: in giver-produced references, and in explainer judgments, which also co-vary with the explainee's gaze. In same-speaker MapTask reference chains, the speaker's gaze entropy is lower at the mention where a previously non-aligned referent becomes aligned. The best gaze feature groups improve modestly over controls under grouped cross-validation: temporal features in MapTask and raw proportions in MUNDEX. Because effects are small and several weaken when recurring participants rather than dialogues are the unit of inference, we treat gaze as one contributing cue to grounding, to be interpreted alongside task and dialogue context.
The Brief
2:33A short podcast summary of the paper. Written and voiced by AI.
Transcript
Gaze as Evidence for Common Grounding. A Cross-Corpus Analysis of MapTask and MUNDEX. By Nan Li and two others, from Utrecht University and Queen Mary University of London. Published on the sixteenth of September. Its code repository is public but slight: five source files.
When two people work together on a task and each knows something the other does not, they have to build a shared understanding as they talk. This is called common grounding. The question here is whether gaze, where each person is looking, gives evidence about that grounding. The answer across two different tasks is a qualified yes. Moments of successful grounding go with more gaze directed at the task, and with gaze toward the partner that is less frequent, less scattered, and less restless.
The two corpora are HCRC MapTask and MUNDEX. Working from discrete annotations of behaviour, the authors map both into one shared vocabulary of gaze: at the partner, at the task, or away. They then compute gaze features around the units of dialogue that matter for the task. In MapTask, the marker of grounding is a reference interpreted in an aligned way. In MUNDEX, it is a judgement of UND, meaning understood.
In both corpora, those grounded moments come with more gaze at the task, less gaze at the partner, lower gaze entropy, and fewer shifts of gaze. The associations are clearest for the participant who leads the task: in references made by the MapTask giver, and in judgements made by the MUNDEX explainer, which also vary with where the explainee is looking. In MapTask chains where the same speaker keeps referring to something, that speaker's gaze entropy drops at the very mention where a referent that had not been aligned becomes aligned.
Used for prediction, the best groups of gaze features improve modestly over controls under grouped cross-validation. Temporal features work best in MapTask, and raw proportions in MUNDEX.
The authors are careful about how far this goes. The effects are small. Several of them weaken when the unit of inference is the recurring participant rather than the dialogue, which means some of the signal may belong to particular people rather than to grounding in general. For that reason they treat gaze as one contributing cue to grounding, not a reliable indicator on its own, to be read alongside the context of the task and of the dialogue.
The value of the comparison is that two quite different tasks, mapped into one vocabulary, tell a consistent story. When understanding is shared, people look at the work more and at each other less, and look around less. It is a modest signal, and the authors present it as one, but it recurs across both settings.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Two corpora, two ways of labelling where people look. First you have to make them speak the same language.
You study how two people build a shared understanding while one guides the other through a task. Two video corpora record where each person looks, but each labels gaze in its own way.
The road ahead
- 1 · The problem ~6 min Two video corpora both record where people look while building shared understanding, but each labels gaze differently, so their findings cannot be compared.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Map both corpora's gaze into partner, task or away, then ask which gaze goes with grounding, and treat it as one modest cue.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Map gaze into a shared vocabulary, compute features around grounding moments, compare grounded with ungrounded moments, then test prediction on held-out groups.
- 6 · The experiment ~6 min You’ll guess: In both corpora, where did people look more during moments of successful grounding?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.