What specific repair mechanisms maintain intersubjectivity during conversation?
This explores the concrete moves people make to keep a shared understanding alive in conversation — and, the corpus shows, exactly which of those moves current AI systems lack or actively erode.
This explores the concrete repair moves that keep two minds aligned mid-conversation — and the collection's sharpest insight is that naming these mechanisms also exposes where AI quietly fails. The clearest inventory comes from work treating conversation maintenance as social action rather than information transfer: humans constantly perform small implicit repairs — re-specifying an ambiguous reference, handing off a topic gracefully, checking that the other person followed — and these sustain the *relationship* of mutual understanding, not the content Why don't language models develop conversation maintenance skills?. A second strand gives these moves a name and a count: "grounding acts," the clarifying questions and understanding-checks that establish shared footing. Models produce 77.5% fewer of them than people do Does preference optimization damage conversational grounding in large language models?, and the cause is diagnostic — preference optimization rewards confident, fluent single-turn answers, so the very training that makes a model seem helpful taxes away the repair work multi-turn conversation depends on Does preference optimization harm conversational understanding?.
The most fundamental repair mechanism is symmetric updating of common ground: when you contradict me or pivot the topic, I absorb that revision into our shared background and we both carry it forward. Here the corpus finds a structural wall — LLMs interpret every later turn inside the frame of the initial prompt and cannot jointly update the scoreboard, leaving the *user* as the sole maintainer of common ground Can LLMs truly update shared conversational common ground?. Asking a clarifying question is itself a repair move, and one model deliberately re-trains for it: rewarding long-term interaction value instead of next-turn helpfulness lets a system actively discover intent rather than passively guess at it Why do language models respond passively instead of asking clarifying questions?.
What's worth knowing that you might not have expected: a few notes supply the formal machinery these repairs would need. Collaborative Rational Speech Acts extends pragmatic reasoning so a system tracks *both* speakers' beliefs across turns and models the progression from partial to shared understanding — the information-theoretic backbone token-level models lack Can dialogue systems track both speakers' beliefs across turns?. And the repair work shows up structurally: explanations succeed not by clean delivery but by co-construction, where topic relation, dialogue act, and explanation move interact turn by turn What makes explanations work in real conversation?. Strikingly, *how* people repair predicts whether a conversation lands almost as well as *what* they say — structural trajectory alone forecasts dialogue satisfaction at near-content accuracy Can conversation structure predict dialogue success better than content?.
The collection's quiet verdict, then, is that intersubjectivity is maintained by a repertoire of mostly-invisible relational moves — reference repair, topic hand-off, grounding checks, clarifying questions, and symmetric common-ground updates — and that today's models either don't perform them or are trained out of them. If you want the deepest cut, one note argues the gap is partly architectural rather than fixable by tuning: humans carry relational continuity in a biological substrate that persists between encounters, while a model is rebuilt from stored text each session, so there's no carrier for the accumulated repair to live in Does an LLM have anything that persists between conversations?.
Sources 9 notes
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.
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.
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.
CRSA integrates rate-distortion theory with RSA to enable bidirectional belief tracking across dialogue turns. Demonstrated on referential games and doctor-patient dialogues, it captures progression from partial to shared understanding, providing the information-theoretic framework that token-level LLM systems lack.
Analysis of 399 daily-life explanations shows that topic relation, dialogue act, and explanation move jointly predict understanding success. Explanations are co-constructed through interaction patterns, not monological delivery—challenging how LLMs currently generate explanations.
TRACE achieved 68% accuracy predicting dialogue success from structural features alone, matching a 70% content-based baseline. A hybrid combining both reached 80%, suggesting how agents communicate rivals what they say.
While humans have a continuous biological-phenomenological substrate that preserves interaction effects during dormancy, LLMs have no analogous carrier. The virtual instance is reconstituted from stored text each time, making resumed and new conversations structurally identical.