What role does accommodation play in making discourse coherent?
This reads 'accommodation' as the mutual adjustment speakers make to each other — checking, repairing, and updating shared understanding — and asks whether that adjustment is what actually holds a conversation together.
This explores accommodation as the ongoing work of speakers adapting to one another — and the corpus suggests coherence isn't a property of well-formed sentences but of that adaptive work being done at all. The clearest statement of what coherence requires comes from work showing it depends on simultaneously tracking three layers: the linguistic segments, the purposes behind them, and what's currently salient in attention How do readers track segments, purposes, and salience together?. Accommodation is how those layers stay aligned across two minds rather than one — when a speaker senses the other has lost the thread, they adjust. Strip that out and coherence degrades into recognizable failure types: contradiction, broken reference, irrelevance, and waning engagement What semantic failures break dialogue coherence most realistically?.
The accommodating moves themselves are small and easy to overlook — clarifying questions, checks for understanding, acknowledgments that signal 'I'm with you.' These 'grounding acts' are the connective tissue, and the striking finding is what happens when you optimize them away. Preference-tuned models produce 77.5% fewer grounding acts than humans, because the training reward favors fluent, confident answers over the visibly tentative work of establishing shared footing Does preference optimization damage conversational grounding in large language models?. The result is an 'alignment tax': models that look helpful turn by turn but fail silently in longer exchanges, because they've stopped accommodating Does preference optimization harm conversational understanding?. Similarly, next-turn reward optimization trains models to respond passively instead of probing for what the user actually meant Why do language models respond passively instead of asking clarifying questions?.
There's a deeper structural reason accommodation matters: coherence is co-constructed, not delivered. Explanations that succeed do so through the interaction itself — topic relation, dialogue act, and explanation move jointly predict whether the listener understands, which means a 'good explanation' isn't a finished object but something built between two people What makes explanations work in real conversation?. This is also why some systems can't accommodate even when they try: LLMs tend to read every later turn through the frame of the opening prompt, so they can't symmetrically update common ground — when the user pivots, the burden of maintaining the shared 'scoreboard' falls entirely on the human Can LLMs truly update shared conversational common ground?.
The surprising turn is that accommodation can be too strong. In collaborative reasoning, frontier models that solve problems alone collapse when working together, reaching over 90% agreement regardless of whether the answer is right — they accommodate so eagerly they lose the ability to productively disagree Why do language models fail at collaborative reasoning?. So coherence isn't maximal agreement; it's calibrated mutual adjustment, including the adjustment of pushing back. Frameworks like collaborative rational speech acts try to formalize this as bidirectional belief tracking — modeling both speakers' evolving understanding rather than just generating the next plausible token Can dialogue systems track both speakers' beliefs across turns?.
What you might not expect: accommodation leaves a measurable trace in conversational *shape*. Structural features of how a dialogue moves predict satisfaction at 68% accuracy — nearly matching what's said — suggesting the rhythm of adjustment itself, not just the content, is what readers register as a conversation that 'worked' Can conversation structure predict dialogue success better than content?.
Sources 10 notes
Discourse processing demands parallel recognition of linguistic segments, intentional structure, and attentional salience—not sequential processing. These three layers constrain each other during comprehension, and failures in any single layer disrupt overall understanding.
Research using Abstract Meaning Representation identified four distinct incoherence types: contradiction, coreference inconsistency, irrelevancy, and decreased engagement. AMR-trained classifiers detect these semantic failures while text-level manipulations alone cannot.
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.
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.
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.
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.
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.
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.
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.