How do dialogue coherence failures map onto the three discourse components?
This explores how the many ways AI dialogue breaks down sort onto the three layers one strand of the corpus says coherence actually runs on — the linguistic segments, the speakers' purposes, and what's currently in focus.
This explores how the various dialogue-breakdown findings in the collection map onto a specific three-part model of discourse: the claim in How do readers track segments, purposes, and salience together? that coherence isn't one thing but three layers tracked at once — the linguistic segments (what was literally said), the intentional structure (what each turn is *for*), and the attentional salience (what's currently in focus). Once you have that frame, the corpus's scattered failure modes stop looking like a grab-bag and start looking like damage localized to one layer or another.
The surface layer — segments — is where the most legible failures live. The AMR work in What semantic failures break dialogue coherence most realistically? names four: contradiction, coreference inconsistency, irrelevancy, and decreased engagement. These are detectable in the linguistic structure itself, which is why a meaning-representation classifier can catch them while plain text-level checks can't. But notice how shallow this layer is — it tells you the conversation has gone incoherent without telling you why.
The *why* almost all sits in the second layer, intentional structure, and this is where the collection is richest. The multi-turn collapse documented in Why do language models fail in gradually revealed conversations? is an intent failure: models lock onto an early guess about what you want and can't recover. Why do AI conversations reliably break down after multiple turns? reframes the whole multi-turn breakdown as intent misalignment rather than a capability ceiling. Why do language models avoid correcting false user claims? shows models declining to correct false claims they actually know are false — a purpose-layer failure driven by social mimicry, not ignorance. And several notes trace this back to training: Why do dialogue failures persist despite scaling language models? argues the repair and common-ground operations that *are* the intentional layer were simply never in the training data, while Does preference optimization harm conversational understanding? and Why do language models respond passively instead of asking clarifying questions? show RLHF actively eroding them — rewarding confident single-turn answers over the clarifying questions that build shared intent.
The third layer, salience — what's in focus right now — is the least directly studied but quietly present. Selective memory retrieval, floated as a fix in Why do AI conversations reliably break down after multiple turns?, is essentially a salience-management move: keep the relevant thing foregrounded. Can tracking dialogue dimensions simultaneously reveal hidden conversation patterns? and Can conversation structure predict dialogue success better than content? both track the trajectory of a conversation over time — topic coherence, emotional arc, structural shape — which is salience viewed as something that moves. The striking finding in the latter is that *how* a conversation is structured predicts satisfaction nearly as well as *what* was said, which is exactly what you'd expect if the attentional layer were doing real work.
The payoff of laying it out this way: the corpus has a lopsided map. Segment-level failures are well-instrumented but shallow, intentional failures are deep and increasingly traced to training and reward design, and salience is the under-explored frontier. And Can dialogue systems track both speakers' beliefs across turns? hints at what a unified treatment looks like — an information-theoretic model that tracks both speakers' beliefs across turns, i.e. the intentional and attentional layers together — precisely the bidirectional bookkeeping token-level LLMs lack.
Sources 11 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.
Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.
Research shows AI conversations degrade due to intent understanding gaps rather than inherent capability deficits. Architectural patterns like mediator-assistant structures and selective memory retrieval recover lost performance without retraining.
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.
LLMs trained on monological written text lack dialogue-specific operations like repair and common-ground construction. Dialogue failures—topic drift, presumption of shared context, absent repair—are absences in the training mode, not capability deficits, and cannot be fixed by scaling text alone.
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.
Conversational DNA encodes four simultaneous dimensions—linguistic complexity, emotional trajectories, topic coherence, and conversational relevance—as temporal streams. The reverse Turing test finding showed expert assessments of AI diverged sharply, suggesting conversational structure shapes interpretation as much as content.
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.
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.