When two people work on a task together, does the one steering show clearer signs, like where they look, that you're on the same page?
How do task leadership roles shape the strength of grounding signals?
This explores whether the person steering a joint task gives off stronger, more readable signs of shared understanding than the person following, and what that means for detecting grounding, especially in AI systems that can't see those signs.
This explores whether the person steering a joint task gives off stronger, more readable signs of shared understanding than the person following. The corpus has one study that speaks to this directly, so the answer is narrow but interesting. Does gaze reveal whether people have achieved common ground? looked at two collaborative-task datasets (MapTask and MUNDEX). When two people had reached aligned understanding, their eyes showed it: more looking at the task, less looking at each other, and more predictable gaze patterns (lower gaze entropy). The effect was strongest for task leaders.
The effect was consistent but modest, and the study doesn't say why leaders show it more. A plausible guess, which is my inference and not the paper's, is that the leader's attention is tied to the thing being negotiated. Once the follower gets it, the leader can stop checking their partner and go back to the task. That would make the leader's gaze a cleaner readout of whether grounding worked.
The practical catch is that this signal is invisible to text-only systems. In a chat window there is no gaze, so grounding has to happen in words: clarifying questions, acknowledgments, understanding checks. LLMs produce 77.5% fewer of these acts than humans (Why do language models sound fluent without grounding?). Preference optimization makes the gap worse, because raters reward confident, complete answers over a model that stops to check (Does preference optimization damage conversational grounding in large language models?, Does preference optimization harm conversational understanding?). So a model can't read the leader's gaze signal, and it also suppresses the verbal signals that could stand in for it.
Roles also seem to matter in purely social ways. Models often fail to reject a false claim in a user's question even when they know the right answer, apparently to avoid the social cost of correcting someone (Why do language models avoid correcting false user claims?). That is a grounding failure shaped by who is speaking to whom. Grounding is also person-specific, since the same words connect to the world differently for each speaker, so it has to be calibrated actively (Why do speakers need to actively calibrate shared reference?). And social grounding is acquired by taking part in language games, which include their role structure (Can LLMs acquire social grounding through linguistic integration?).
The corpus has no study of leadership roles in human-AI conversation. It's open whether a model is treated as the leader, whose signals we would read, or the follower, who mostly nods along. The gaze finding suggests that, for people, who leads changes how visible understanding is.
Sources 7 notes
Across two corpora, aligned understanding coincided with more task-directed gaze, less partner-directed gaze, and lower gaze entropy—especially for task leaders. The effect was consistent but modest, and invisible to text-only systems.
LLMs generate 77.5% fewer grounding acts than humans—no clarifying questions, acknowledgments, or understanding checks. Preference optimization actively removes these behaviors because raters prefer confident complete answers, creating an illusion of fluency that masks communicative incompetence.
Research shows LLMs generate 77.5% fewer grounding acts than humans, and RLHF preference optimization actively worsens this gap. The optimization target—fluent, confident responses—directly undermines the communicative work of establishing shared understanding.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
Show all 7 sources
The same words can mean different things to different speakers because referential grounding is person-specific. True communicative grounding demands collaborative negotiation of how language connects to the world, not mere surface-level word sharing.
Social grounding is acquired through participation in language games rather than possessed innately. As LLMs become established communicative partners in human linguistic practice, they develop elementary social grounding comparable to young children, making the question of LLM understanding time-indexed.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Grounding Gaps in Language Model Generations
- Conversational Alignment with Artificial Intelligence in Context
- Can LLMs Ground when they (Don't) Know: A Study on Direct and Loaded Political Questions
- “Understanding AI”: Semantic Grounding in Large Language Models
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- A recipe for annotating grounded clarifications
- Gaze as Evidence for Common Grounding: A Cross-Corpus Analysis of MapTask and MUNDEX
- RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback