Can static word-sharing create genuine communicative grounding between humans and models?
This explores whether sharing the same vocabulary is enough to produce real mutual understanding between humans and LLMs — or whether grounding is a dynamic, two-sided process that static word-matching can never reach.
This explores whether sharing the same vocabulary is enough to produce genuine grounding — and the corpus answers, fairly bluntly, no: words are the visible residue of grounding, not the thing itself. The clearest statement of the problem is that LLMs treat the opening prompt as a fixed frame and can't jointly update common ground Can LLMs truly update shared conversational common ground?. When two humans talk, both partners keep revising a shared scoreboard of what's mutually assumed; with a model, the user ends up maintaining that scoreboard alone. So even when both sides use identical words, only one side is actually negotiating their meaning.
There's a deeper version of this. Bender & Koller's argument is that meaning lives in the relation between expressions and communicative intent, and a system trained on form-to-form prediction with no access to shared attention simply has no channel to that relation Can language models learn meaning from text patterns alone?. A complementary framing says AI output is 'event-residue' — it carries the surface markers of communication inherited from training, but lacks the event structure of a real utterance, so the human supplies the missing orientation through interpretive labor Does AI generate genuine utterances or just text patterns?. Put together: static word-sharing creates the *appearance* of grounding precisely because humans are so good at animating dead text into a felt exchange.
What makes this hard to notice is that the fluency is partly an artifact of skipped work. Models produce ~77% fewer grounding acts than humans — fewer clarifying questions, acknowledgments, understanding checks — and preference optimization actively strips these out because raters reward confident, complete answers Why do language models sound fluent without grounding?. The maintenance moves that actually build common ground are social actions, not information transfer, so a training signal that rewards prediction never grows them Why don't language models develop conversation maintenance skills?. Models even avoid correcting false claims to save face, mirroring human politeness norms without the underlying commitment to a shared truth Why do language models avoid correcting false user claims?.
The more interesting part of the corpus is where it points past mere word-sharing toward something more dynamic. Lexical entrainment — drifting toward your partner's word choices — is largely absent from current systems, but can be partly taught through preference tuning on coreference signals Why don't conversational AI systems mirror their users' word choices?. Training models to ask genuinely useful clarifying questions Can models learn to ask genuinely useful clarifying questions? and rewarding long-horizon interaction value rather than next-turn helpfulness Why do language models respond passively instead of asking clarifying questions? both push toward active intent discovery — the two-sided revision that static frames lack. And there's a wilder horizon: sharing latent thoughts directly between agents rather than through words at all, detecting alignment conflicts at the representational level Can agents share thoughts directly without using language?.
The thread worth leaving with: grounding in these notes is never about overlapping vocabulary — it's about whether both parties can revise what they jointly assume. That reframes a lot of perceived AI 'understanding' as one-sided interpretive work the human is doing for free, and it suggests the real engineering target isn't better word-matching but giving models a stake in the shared scoreboard.
Sources 10 notes
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.
Bender & Koller argue that meaning requires the relation between expressions and communicative intents. Since LLMs are trained only on form-to-form prediction with no access to shared attention or intent, they cannot reconstruct the meaning that grounds language.
AI output carries communicative markers inherited from training data but lacks the event structure that produces actual utterances. Users supply the missing orientation through interpretive labor, creating a pseudo-event with structure only on the human side.
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.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
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.
Response generation models fail to adapt vocabulary toward users' lexical choices, a phenomenon central to human rapport and clarity. Post-training via DPO on coreference-identified preferences can teach models in-context convention formation.
The ALFA framework breaks down question quality into theory-grounded attributes (clarity, relevance, specificity) and trains models on 80K attribute-specific preference pairs. Attribute-specific optimization outperforms single-score training, especially in clinical reasoning where asking the right clarifying question directly impacts decision quality.
CollabLLM demonstrates that standard RLHF training optimizes for immediate helpfulness, discouraging models from asking clarifying questions or offering multi-turn insights. Multi-turn-aware rewards that estimate long-term interaction value enable active intent discovery and genuine collaboration.
Research formalizes inter-agent thought sharing via sparse autoencoders that recover individual, shared, and private latent thoughts from hidden states. This approach detects alignment conflicts at the representational level before they manifest in language.