Can convention formation improve communicative grounding beyond word sharing?
This explores whether the back-and-forth of building shared conventions — not just using the same words — is what actually creates mutual understanding, and whether LLMs can be taught to do it.
This explores whether the back-and-forth of building shared conventions — not just using the same words — is what actually creates mutual understanding, and whether LLMs can be taught to do it. The corpus is emphatic on the first half: word sharing is necessary but nowhere near sufficient. The same word can point to different things for different speakers, so real grounding demands ongoing calibration of how language hooks onto the world, negotiated collaboratively rather than assumed Why do speakers need to actively calibrate shared reference?. Convention formation is precisely that calibration in motion — partners gradually settling on shortened, shared ways of referring to things they've established together.
The most direct evidence that this is teachable comes from work that post-trains models to form ad-hoc conventions in real time: using preference pairs that reward shortening references once they've been mutually established (but punish shortening them prematurely), plus special tokens that mark re-mentions, models begin spontaneously forming conventions during a conversation without task-specific tuning Can we teach LLMs to form linguistic conventions in context?. That's grounding beyond word sharing in action — the meaning of a compressed reference depends on shared history, not the dictionary.
But here's the twist the corpus keeps circling: the dominant training regime works against this. Preference optimization rewards confident, complete, single-turn answers, and in doing so strips out the clarifying questions, acknowledgments, and understanding-checks that humans use to ground — models produce 77.5% fewer of these grounding acts than people Why do language models sound fluent without grounding? Does preference optimization damage conversational grounding in large language models?. Fluency becomes a mask for communicative incompetence, an "alignment tax" that looks helpful while quietly failing in multi-turn exchanges Does preference optimization harm conversational understanding?. So convention formation and standard alignment pull in opposite directions.
There's a deeper architectural obstacle too. Convention formation assumes both partners can update the shared scoreboard, but LLMs tend to treat the opening prompt as a fixed frame and interpret everything afterward inside it — they can't symmetrically absorb revisions into jointly held common ground, leaving the user as the sole bookkeeper Can LLMs truly update shared conversational common ground?. Alignment also locks models into one static communicative identity, blocking the register-switching that real pragmatic negotiation requires Can language models adapt communication style to different contexts?. The optimistic counterweight is that grounding may not be innate but acquired through participation: as LLMs become established communicative partners in human practice, they can develop elementary social grounding the way young children do Can LLMs acquire social grounding through linguistic integration?.
The thing you didn't know you wanted to know: convention formation isn't a nicety layered on top of communication — it may be the load-bearing mechanism, and the very training that makes models sound articulate is the thing sabotaging their ability to do it.
Sources 8 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.
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.
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 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.
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.