Why does shared practice matter for meaning to take hold?
This explores why meaning isn't something a word carries on its own, but something that takes hold only when speakers do shared communicative work together — and what the corpus shows happens to machines (and people) when that shared practice is missing.
This explores why meaning isn't something a word carries on its own, but something that takes hold through shared practice — and the corpus is surprisingly unified on the answer: meaning is *negotiated*, not transmitted. The same words can mean different things to different speakers because grounding is person-specific; understanding each other requires actively calibrating what a word points to in the world, not just exchanging the word itself Why do speakers need to actively calibrate shared reference?. So shared practice matters because reference is collaborative from the start — meaning lives in the back-and-forth where two people confirm they're talking about the same thing.
A striking thread is that this view treats understanding as something you *acquire by participating*, not something you're born with. Social grounding develops through taking part in language games, which is why LLMs gain elementary social grounding as they become established communicative partners — making the question "does the model understand?" time-indexed rather than yes/no Can LLMs acquire social grounding through linguistic integration?. That same paper's framework splits grounding into functional, social, and causal kinds, with LLMs strong on function but weak (though growing) on the social dimension that practice supplies Does semantic grounding in language models come in degrees?. But participation only goes so far: even heavy use can't grant linguistic agency in the full sense, which the corpus argues requires embodiment and stakes no amount of practice provides Do LLMs gain true linguistic agency through integration?.
Where shared practice is most visibly *absent*, you can watch meaning fail to take hold. LLMs treat the opening prompt as a fixed frame and can't jointly update the shared scoreboard of a conversation — when you pivot or contradict an earlier assumption, the model can't absorb the revision, so the human ends up the sole maintainer of common ground Can LLMs truly update shared conversational common ground?. And the very training that makes models feel helpful actively erodes the small repair acts — clarifying questions, checks of understanding — that grounding depends on; preference optimization rewards confident single answers and cuts grounding behaviors by over 75% below human levels Does preference optimization harm conversational understanding? Does preference optimization damage conversational grounding in large language models?. Models that ace problems alone collapse when they have to reason *with* a partner, agreeing over 90% of the time regardless of who's right Why do language models fail at collaborative reasoning?.
The payoff — the thing you might not have known you wanted to know — is that meaning-making can be *taught back in*, and it shows up as a kind of social competence rather than raw knowledge. Train models to stay consistent only when a partner's intervention genuinely matters, and partner-awareness emerges as a byproduct without any explicit reward for it Why do standard alignment methods ignore partner interventions?. Push further and models will spontaneously invent ad-hoc conventions mid-conversation — shortening how they refer to something once it's mutually established — exactly the way two people coin private shorthand Can we teach LLMs to form linguistic conventions in context?. That's shared practice doing its work in miniature: a convention is meaning that took hold because two parties used it together.
The shadow side is worth knowing too. Shared background isn't always built honestly — presuppositions persuade better than direct claims precisely because they smuggle new content in as though it were already common ground, bypassing scrutiny Why are presuppositions more persuasive than direct assertions?. And in human practice, the texture of grounding is subtle: therapists who lean on "I" build weaker alliances, while a patient's filler-laden pauses can signal the relaxed, trusting footing where shared meaning actually settles Does therapist self-reference language predict weaker therapeutic alliance?. Across all of it, the lesson holds — meaning takes hold in the practice, not in the words.
Sources 12 notes
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.
Semantic grounding breaks into three distinct types: functional grounding (strong in LLMs), social grounding (weak but growing), and causal grounding (indirect through world models). LLMs score differently on each dimension, making the yes-or-no understanding question misleading.
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 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.
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.
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.
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.
Regularizing agents to maintain consistency when intervention pathways are nullified forces them to evaluate suggestions by causal impact rather than surface plausibility. Common ground alignment emerges as a byproduct without explicit reward.
Post-training with two types of preference pairs derived from TV scripts — one encouraging re-mention shortening, one preventing premature shortening — plus special [remention] tokens enables models to spontaneously form ad-hoc linguistic conventions during interaction without task-specific fine-tuning.
Experimental evidence shows presuppositions with additive, iterative, and factive triggers persuade audiences more than assertions, especially for discourse-new content. The mechanism: presuppositions bypass evaluative scrutiny by presenting claims as already-accepted background.
High frequency of therapist 'I' usage correlates with lower patient-reported alliance and reduced trusting behavior in validated behavioral tasks. Patient non-fluency markers like filler pauses, conversely, signal relaxed communication and stronger alliance.