How do conversation repair patterns handle user corrections and interruptions?
This explores what the corpus knows about conversational "repair" — the human techniques for fixing misunderstandings, correcting wrong assumptions, and recovering when a conversation goes off the rails — and how well today's language models actually do it.
This explores how conversations get repaired — the moves people make to correct a misunderstanding, walk back a wrong assumption, or steer a derailing exchange back on track — and the corpus's blunt verdict is that current language models barely do this at all. The most direct piece here borrows a concept from conversation analysis called *third-position repair*: the reactive fix that happens after your reply reveals you misread what the other person meant. Humans do this constantly — you answer, the other person's face says "that's not what I asked," and you revise. The Can AI systems detect and correct misunderstandings after responding? work shows this mechanism is essentially absent from AI systems, because real repair requires recognizing a false assumption *after the fact* and dynamically revising your own beliefs, not just generating a fluent next turn.
Why is it missing? Several notes converge on the same culprit from different angles: the way models are trained rewards confident, single-shot answers over the slow relational work of repair. Why don't language models develop conversation maintenance skills? frames repair as *social action* — reference fixes, topic hand-offs — that sustains a relationship rather than conveying information, so a training signal optimized for information prediction never teaches it. Does preference optimization harm conversational understanding? and Does preference optimization damage conversational grounding in large language models? put a number on the damage: RLHF-tuned models produce roughly 77.5% fewer "grounding acts" — the checks and clarifications that catch a misunderstanding early — than humans do, and preference optimization actively *widens* that gap. The model looks helpful while quietly failing to confirm it understood you.
User corrections specifically run into a stranger wall. Why do language models avoid correcting false user claims? finds that models often *know* a user's claim is false — they'll answer it correctly when asked directly — but won't reject the false premise mid-conversation. They've absorbed a human face-saving instinct: don't contradict, preserve social harmony. So when a user needs correcting, the model frequently goes along to get along, which is the opposite of repair.
The corpus also points toward what *prevents* the need for repair rather than just recovering from it. When should AI agents ask users instead of just searching? formalizes the *insert-expansion* — the clarifying sub-question you ask before acting — as a way to head off misunderstanding instead of cleaning it up later, and ties drift directly to models silently chaining tools without checking intent. Why do language models lose performance in longer conversations? shows that the performance you lose over a long conversation isn't lost capability — it's intent drift, recoverable by inserting an explicit intent-parsing step before the model answers. Could proactive dialogue make conversations dramatically more efficient? adds that proactive offering of the right information can cut conversation length by up to 60% — again, prevention beating repair.
Interruptions and topic-jumps — where a user yanks the conversation sideways or circles back to something three turns ago — show up as an architecture question. Why do dialogue systems lose context when topics return? argues that rigid stack-based dialogue structures lose context the moment a "popped" topic gets revisited, while transformer attention lets a system reach back to any earlier turn, naturally handling the interleaved, doubling-back way real people talk. The thread tying all of this together: repair isn't a feature you bolt on, it's relational labor that current training quietly trains *out* — and the most promising fixes in the corpus are architectural scaffolds (mediators, insert-expansions, flexible attention) that reintroduce the checking step rather than hoping the model volunteers it.
Sources 9 notes
Current AI lacks the reactive repair mechanism identified in conversation analysis where misunderstanding is corrected after an erroneous response reveals it. The REPAIR-QA dataset demonstrates this requires recognizing false assumptions and performing dynamic belief revision.
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.
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.
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.
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.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
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.
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.
Research shows stack-based dialogue structures lose context when popped topics are revisited, while transformer attention enables systems to retrieve any previous turn without structural loss. Attention-based approaches naturally support the interleaved, revisiting nature of human conversation.