Why do language models use twice as many words per conversation turn?
This reads the question as being about verbosity — why model turns balloon with words — and treats that bloat not as a quirk of formatting but as a downstream symptom of how these systems are trained to behave in conversation.
This explores why conversational turns from language models tend to run long, and the corpus doesn't have a note that literally counts words — but it has a strong explanation for the behavior underneath the word count. The short version: models are trained to answer rather than to ask, and when they don't know exactly what you mean, the cheapest way to look helpful is to say more. RLHF rewards immediate, complete-feeling responses over clarification-seeking, so a model facing an underspecified request doesn't pause to narrow things down — it front-loads an answer that tries to cover every reading at once Why do language models lose performance in longer conversations? Why do language models respond passively instead of asking clarifying questions?. That hedging-across-interpretations is exactly what doubles the word count.
There's a deeper mechanism here worth knowing about. When information arrives gradually across a conversation, models lock into an early guess and can't course-correct — accuracy drops sharply (around 39% across 200,000+ conversations) once a single instruction is spread over multiple turns Why do language models fail in gradually revealed conversations? Why do AI assistants get worse at longer conversations?. A model that has committed to a premature assumption tends to overexplain it: justifying, caveating, and elaborating the guess instead of just asking which thing you meant. So verbosity and the well-documented "getting lost" failure are two faces of the same training incentive — one rewards talking, none rewards checking.
The more surprising angle the corpus offers is what's *missing* that would otherwise keep turns short. Human conversation stays lean through implicit social moves — repairing a reference, handing off a topic, mirroring the other person's vocabulary — that do relational work rather than transmit information Why don't language models develop conversation maintenance skills?. Models never learn these because the training signal rewards predicting information, not sustaining a relationship. One concrete example: humans naturally converge on each other's word choices (lexical entrainment), which compresses dialogue over time; current systems simply don't do this and instead re-explain from scratch each turn Why don't conversational AI systems mirror their users' word choices?.
If you want to go deeper, the interesting thread is that none of this looks like a capacity limit — it's an absent training objective. Models can be taught what to ignore and when to stay terse with surprisingly little data; one study fixed topic-drift resilience with just over a thousand synthetic dialogues Why do language models engage with conversational distractors?. The takeaway you didn't know you wanted: the wordiness isn't the model being thorough — it's the model substituting volume for the clarifying question it was trained not to ask.
Sources 7 notes
LLMs degrade in multi-turn settings because RLHF training rewards premature answers over clarification-seeking, creating pragmatic mismatch with individual user behaviors. A Mediator-Assistant architecture that explicitly parses user intent before execution recovers lost performance without retraining.
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.
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.
LLMs perform at 90% accuracy with single-message instructions but drop to 65% across natural conversation. Models lock into early guesses when information arrives gradually and cannot course-correct, a behavior induced by RLHF training that rewards helpfulness over clarification.
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.
Fine-tuning on just 1,080 synthetic dialogues with distractor turns significantly improves topic resilience, revealing that the gap is not model capacity but absent training signal. Models learn to follow what-to-do instructions but not what-to-ignore instructions.