AI assistants rarely pause to check they've understood you right — they just answer with total confidence anyway.
Do LLMs presume common ground without actively building it with partners?
This explores whether LLMs check that they and the person they're talking to actually understand each other, or whether they just act as if they already do.
This explores whether LLMs work to build shared understanding with the person they're talking to, or simply assume it's already there. The corpus says the second, and gives a number: LLMs produce grounding acts (clarifying questions, acknowledgments, repairs like "wait, did you mean…?") 77.5% less often than humans do Do language models actually build shared understanding in conversation?. Instead of checking in, they give fluent answers in a confident tone. That confidence is the subtle part. An authoritative response hides the fact that no checking happened, so the reader rarely notices that their meaning was never confirmed.
The problem runs deeper than rarely asking questions. Common ground is the running record of what both people in a conversation take for granted, and each turn should be able to update it. One line of analysis argues that LLMs can't really update this record together with the user at all. They read every later turn through the frame set up at the start Can LLMs truly update shared conversational common ground?. If you change topic or contradict your earlier framing, the model doesn't fold that change into a shared background. In practice this makes the user the only one keeping score. If you've ever felt you had to re-explain your situation to a chatbot halfway through, this is probably why.
A related pattern shows up when LLMs talk to each other instead of to people. Agents in multi-agent teams commit early to partners without exploring what those partners know or can do, and they need structured prompting before they'll look around Why don't LLM agents naturally explore each other in teams?. When frontier models reason together, they agree with each other over 90% of the time whether or not the answer is right. As a result, group performance falls below what each model manages alone Why do language models fail at collaborative reasoning?. This is the same habit seen from another side: going along with what seems shared instead of testing it. The good news is that training for productive disagreement improved results by 16.7%, so this looks more like a learnable social skill than a fixed limit. A contrasting result: reasoning models given direct access to each other's working memory (a shared KV cache) coordinate on their own, without being told to Can multiple LLMs coordinate without explicit collaboration rules?. Grounding may come easier when partners can see each other's thinking directly than when they have to negotiate it in conversation.
The philosophical notes add a time dimension. One argument holds that social grounding isn't something a model either has or lacks. It grows as LLMs take part in human language practices, and current models may sit at roughly a young child's level Can LLMs acquire social grounding through linguistic integration?. A companion note sets a ceiling on this: models can gain more social grounding, but they can't become full conversational agents in the deeper sense, because that requires a body and real stakes Do LLMs gain true linguistic agency through integration?. The same rigidity shows up in values. A model's tone and refusals follow fixed defaults set during training rather than being worked out with the person in context Can language models balance competing ethical norms in context?.
The takeaway you might not have expected: the gap isn't mainly about intelligence. It's about conversational habits that most training doesn't reward, like asking, confirming, and pushing back. Until models are trained to do these things, the practical burden falls on you. Restate your goals when they change, and treat a confident answer as a claim about what the model assumes you meant.
Sources 8 notes
LLMs produce grounding acts—clarifications, acknowledgments, repairs—77.5% less frequently than humans. They generate fluent responses without verifying shared understanding, relying instead on authoritative framing that masks the absence of genuine communicative calibration.
LLMs interpret all subsequent conversational turns within a fixed initial prompt frame, preventing them from symmetrically proposing updates to shared assumptions. Even when users pivot topics or contradict earlier framings, the model cannot absorb revisions into jointly held background—making the user the sole maintainer of conversational scoreboard.
LLM agents exhibit myopic, polarized interaction and commit prematurely to peers without adequate exploration. Structured peer selection (MACE) explicitly induces exploration and improves performance, with gains increasing as agent diversity grows.
Frontier LLMs that solve problems alone fail when collaborating, achieving >90% agreement regardless of correctness. Self-play preference training improves outcomes by 16.7%, suggesting social skills for effective disagreement can be trained.
Existing reasoning-capable models like QwQ and DeepSeek-R1 spontaneously formulate plans, detect redundancy, and adapt strategies when given shared access to a concurrent KV cache. This coordination emerges without fine-tuning, suggesting reasoning models already possess multi-agent collaboration capabilities.
Show all 8 sources
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.
Social grounding and linguistic agency are distinct properties. LLMs acquire more social grounding through integration into language communities, but remain categorically incapable of linguistic agency in the enactive sense, which requires embodiment and precariousness no amount of use can provide.
LLMs cannot perform the situated trade-offs that human pragmatic competence requires. Their ethical principles are structural defaults set at training time, not negotiable moves adapted to context, creating a gap between ethical adherence and communicative appropriateness.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Conversational Alignment with Artificial Intelligence in Context
- Can LLMs Ground when they (Don't) Know: A Study on Direct and Loaded Political Questions
- Grounding Gaps in Language Model Generations
- Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm
- Assertion, Accountability, and Large Language Models
- “Understanding AI”: Semantic Grounding in Large Language Models
- ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey