How does entrainment between speaker and listener build mutual scaling?
This reads 'entrainment' as the way conversation partners drift toward each other's word choices and rhythms, and 'mutual scaling' as the question of whether that mutual adaptation compounds into shared understanding — and the corpus's sharpest point is that humans do this reciprocally while today's AI mostly doesn't.
This explores entrainment — the way a speaker and listener gradually adopt each other's vocabulary and style — and whether that back-and-forth actually builds something larger than either turn alone. In human dialogue it clearly does: when two people converge on the same words for the same things, they're not just being polite, they're constructing a shared reference frame that makes every later turn cheaper to understand. The corpus frames this as lexical entrainment, and notes that current conversational AI conspicuously lacks it — models don't mirror a user's word choices even though that mirroring is central to rapport and clarity in human conversation Why don't conversational AI systems mirror their users' word choices?.
The reason entrainment 'scales' is that it's reciprocal calibration, not one-sided copying. Sharing words isn't the same as sharing meaning — the same word can point at different things for different people, so partners have to actively negotiate how language hooks onto the world Why do speakers need to actively calibrate shared reference?. Each successful round of that negotiation raises the floor for the next, which is the 'mutual' part of mutual scaling. This is exactly where LLMs break: they treat the opening prompt as a fixed frame and interpret every later turn inside it, so they can't symmetrically propose updates to the shared ground. The user ends up as the sole keeper of the conversational scoreboard, doing all the calibration alone Can LLMs truly update shared conversational common ground?.
Here's the part you might not expect: alignment isn't one thing, and the kind of entrainment you build depends on which channel you're matching. A 2020–2025 review found that lexical alignment drives task efficiency and comprehension, while emotional and prosodic alignment drive warmth and trust — and conflating them produces category errors like a chatbot that's word-accurate but cold Do different types of alignment serve different conversational goals?. So 'mutual scaling' isn't a single dial; matching vocabulary scales understanding, matching tone scales relationship, and they don't substitute for each other.
There's also a darker mirror. Entrainment isn't automatically benign coordination: linguistic style matching actually *increases* during deception, because a motivated speaker and listener coordinate language more tightly when something false is being communicated — which means the listener's adaptive mirroring becomes a detectable signal, not just the liar's words Do liars and listeners coordinate their language during deception?. Coordination scales whatever's being coordinated, truth or otherwise.
The corpus's most useful warning is that the way we currently train models actively erodes the machinery entrainment needs. Optimizing for single-turn helpfulness rewards confident answers over clarifying questions and understanding-checks, cutting grounding acts to roughly a fifth of human levels — an 'alignment tax' where the model looks helpful but quietly stops doing the reciprocal work that lets shared meaning accumulate Does preference optimization harm conversational understanding?. And a related pressure runs the other way: rather than the model entraining to the user, the user rephrases toward the high-frequency forms the model handles best, flattening their own distinctiveness on the way in Does high-frequency text homogenize user input before generation?. One cheaper path back toward genuine entrainment is giving the agent an imaginary listener — having it simulate whether its own utterance would actually land for the other party before speaking Can imaginary listeners reduce dialogue agent contradictions?. Mutual scaling, in short, requires both sides to keep adjusting; strip the adjustment from one side and you get accommodation, not entrainment.
Sources 8 notes
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 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.
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.
A 2020–2025 systematic review shows lexical alignment drives task efficiency and comprehension, while emotional and prosodic alignment drive relational warmth and trust. Conflating them in design produces category errors—cold customer-service bots and evasive mental-health assistants.
Research shows interlocutors' linguistic styles correlate more during false communication than truthful communication, especially when the speaker is motivated to deceive. This coordination serves as a detectable deception signal through the listener's adaptive behavior, not just the liar's language.
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.
Adam's Law shows LLMs flatten distinct prompts at comprehension time as users rephrase toward higher-frequency forms the model handles best. The same distributional property that creates accuracy on common tasks filters out distinctiveness on the input side.
Endowing dialogue agents with an imaginary listener via Rational Speech Acts reduces persona contradiction at inference time without NLI labels or extra training. The agent simulates whether utterances would distinguish its persona from a distractor, suppressing generic or contradictory responses.