Why does dialogue-shaped text fail to produce dialogue-like operations in practice?
This explores why text that *looks* like conversation — turns, speakers, back-and-forth — doesn't make a system that actually *does* the things conversation does (tracking who believes what, repairing misunderstandings, adapting to the other person).
This explores why dialogue-shaped text fails to produce dialogue-like operations: the short answer running through the corpus is that real conversation is *social action*, and the format on the page is only its residue. When you train a model to predict the next token in a transcript, you reward it for reproducing what dialogue looks like, not for performing the work dialogue does. Why don't language models develop conversation maintenance skills? makes this the sharpest: humans keep talk smooth through implicit moves — reference repair, topic hand-offs — that sustain a relationship rather than transmit information. Models don't pick these up because the training signal rewards information prediction, not relational work. The maintenance moves leave almost no textual trace to imitate, so imitating the text misses them entirely.
Look at the specific operations that go missing and a pattern emerges: they're all things the surface form doesn't record. Lexical entrainment — drifting toward your partner's word choices to build rapport — is largely absent from conversational AI, because adapting vocabulary mid-conversation is a behavior, not a feature visible in any single turn Why don't conversational AI systems mirror their users' word choices?. Proactivity — volunteering relevant information before being asked, which can cut conversations by up to 60% — is nearly nonexistent in the datasets and benchmarks models learn from, so there's nothing to copy Could proactive dialogue make conversations dramatically more efficient?. And belief tracking across turns, the machinery that lets two speakers move from partial to shared understanding, requires an information-theoretic framework that token-level systems simply don't have Can dialogue systems track both speakers' beliefs across turns?. The dialogue shape is present; the operations underneath it were never in the text to begin with.
This is why models fail in characteristic, non-random ways. Across 200,000+ conversations, every major LLM drops ~39% in multi-turn settings because it locks onto a premature guess early and can't recover — it processed the turns as accumulating text rather than as a belief it should hold loosely and revise Why do language models fail in gradually revealed conversations?. Compare this to how speech-dialogue systems handle the same uncertainty: facing 15–30% recognition error, they maintain *belief distributions* over what the user meant rather than committing to one reading Why do dialogue systems need probabilistic reasoning?. The operation that's missing — staying uncertain, tracking a distribution — is exactly what produces good dialogue and exactly what next-token imitation doesn't install.
There's an even deeper version of this in Do large language models actually commit to a single character?: regenerate the same model response and you get different answers, each consistent with context. The model isn't a participant holding a position; it's sampling from a superposition of possible participants. Dialogue-*like* operation presumes a single agent with commitments to track. The model's relationship to the dialogue shape is performative, not committed — which is why coherence failures (contradiction, broken coreference, irrelevance) show up at the *semantic* level and can't be caught by text-surface manipulation alone What semantic failures break dialogue coherence most realistically?.
The constructive corner of the corpus agrees by inversion: the fixes all stop treating dialogue as text-to-be-completed. Rasa reframes understanding as generating *commands* — pragmatics, what the user wants done — instead of classifying the semantics of what they said Can command generation replace intent classification in dialogue systems?. Other work finds that *how* people talk (structural trajectory) predicts conversational success nearly as well as *what* they say Can conversation structure predict dialogue success better than content?, and that different alignment dimensions serve different goals — lexical for task efficiency, emotional for trust — so flattening them into one undifferentiated text stream produces category errors like the cold service bot Do different types of alignment serve different conversational goals?. The thing you didn't know you wanted to know: some researchers conclude the honest move is to stop pretending the chat transcript is the right interface at all, and generate task-specific UIs instead, which users prefer in over 70% of cases Do generated interfaces outperform text-based chat for most tasks?. If the dialogue shape can't carry the operations, maybe the operations don't belong in dialogue shape.
Sources 12 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.
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.
Simulations show proactivity—providing relevant information without being asked—cuts dialogue turns by 60% in medium-complexity domains. This behavior mirrors human conversation and Grice's maxims but is almost entirely absent from AI datasets and research benchmarks.
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.
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.
Real-world speech recognition achieves 15-30 percent error rates in noisy environments, making deterministic flowchart dialogue systems unworkable. POMDP-based systems handle this by maintaining belief distributions over user intent rather than committing to single interpretations.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
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.
Rasa's dialogue understanding architecture generates domain-specific commands instead of classifying intents, eliminating annotation requirements, handling context naturally, and scaling without degradation—treating understanding as pragmatics rather than semantics.
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.
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 users strongly prefer LLM-generated interactive interfaces—dashboards, tools, animations—over text blocks, especially for structured and information-dense tasks. Structured representation and iterative refinement reduce cognitive load.