Does community integration change LLM properties or only relational positioning?
This explores whether becoming a participant in a language community actually alters what an LLM *is* (its internal capacities) or merely changes where it sits in a web of relationships — and the corpus suggests the answer splits cleanly along the line between two different properties.
This explores whether being woven into human linguistic practice changes an LLM's actual capabilities or only its social standing — and the most useful finding in the corpus is that these come apart. One line of work argues that *social grounding* genuinely accrues through use: grounding isn't something a model is born with, it's earned by participating in language games, so as LLMs become established communicative partners they pick up an elementary, child-like grounding that grows over time Can LLMs acquire social grounding through linguistic integration?. By that account, integration isn't cosmetic — it shifts a real property, and it makes the question of whether a model 'understands' time-indexed rather than fixed.
But the same territory contains a sharp counter-boundary. Social grounding and *linguistic agency* are distinct properties, and only the first is on offer. No amount of community participation grants agency in the enactive sense, because that requires embodiment and precariousness — having something at stake — which use alone can never supply Do LLMs gain true linguistic agency through integration?. So the honest answer to your question is: both, depending on which property you name. Integration moves grounding; it leaves agency categorically untouched.
The 'relational positioning only' reading gets reinforced when you look at what stays frozen even mid-conversation. A model can't jointly update conversational common ground — it reads every later turn through the fixed frame of its initial prompt, so the human ends up being the sole keeper of the shared scoreboard Can LLMs truly update shared conversational common ground?. Relatedly, alignment training locks in a single communicative identity that can't switch register or renegotiate values across contexts, so users can't reshape the model's behavior through dialogue the way they would with a person Can language models adapt communication style to different contexts?. In both cases the model occupies a relational position without the underlying property changing.
Where properties *do* shift, it tends to be through training rather than mere presence in a community. Models that fail at collaborative reasoning — collapsing into >90% agreement regardless of who's right — can be improved 16.7% by self-play preference training, suggesting the social skill of productive disagreement is learnable, not conferred by participation Why do language models fail at collaborative reasoning?. And at the level of architecture, you can inject collaborative-filtering signals directly into the token space to graft on a genuinely new capacity while preserving the old one Can LLMs gain collaborative filtering strength without losing text understanding?. The pattern: real property change comes from intervention on the model, while integration alone mostly rearranges relationships around an unchanged core.
The quiet surprise here is that 'community integration' barely refers to one thing. Putting many LLMs *together* doesn't reliably produce emergent collective competence either — agent groups fail to reach consensus more as the group grows, stalling through timeouts rather than disagreement Can LLM agent groups reliably reach consensus together?, though imposing light external structure while letting agents pick their own roles does outperform both rigid hierarchies and pure autonomy Do self-organizing agent teams outperform rigid hierarchies?. So integration can change *system-level* behavior even when it leaves each model's individual properties exactly where they were.
Sources 8 notes
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 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.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
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.
CoLLM maps traditional collaborative filtering embeddings into the LLM's input token space, letting the LLM attend to CF signals alongside text without modification. This hybrid architecture maintains semantic understanding for cold items while gaining collaborative strength for warm interactions.
Across hundreds of simulations, LLM-agent groups frequently fail to reach valid agreement due to timeouts and stalled convergence rather than subtle value corruption. Agreement degrades with group size even without Byzantine agents present.
A 25,000-task experiment across 8 models and multiple agent counts showed that sequential protocols with external ordering but internal role selection outperform centralized systems by 14% and fully autonomous systems by 44%. Agents spontaneously invented specialized roles and self-abstained when incompetent.