How do dialogue acts and explanation moves interact to predict understanding success?
This explores what makes explanations actually land in conversation — and the corpus reframes it: understanding isn't delivered, it's co-built turn by turn, which is exactly where today's chat models break down.
This explores what makes explanations actually succeed in conversation — and the most direct answer in the corpus is that no single move does the work; understanding emerges from how dialogue acts and explanation moves interact. A study of 399 everyday explanations found that three dimensions — topic relation, dialogue act (what kind of conversational turn you're making), and explanation move (how you're explaining) — jointly predict whether the listener actually gets it What makes explanations work in real conversation?. The headline isn't any one factor but their interaction: explanations are co-constructed, not monologued. A parallel finding in explainable-AI research arrives at the same place from a different angle — explanation quality lives in the source-framing-recipient triad, not in the explanation itself, so who says it and what role the listener plays matter as much as the content What if XAI is fundamentally a communication problem?.
If understanding is co-built, the obvious question is whether LLMs do the co-building — and the corpus says they mostly skip it. Models presume common ground rather than establishing it, producing the clarifications, acknowledgments, and repairs that humans use to check shared understanding about 77.5% less often than people do Do language models actually build shared understanding in conversation?. The reason is partly in the training: preference optimization rewards confident single-turn answers over clarifying questions and understanding checks, which silently erodes exactly the grounding acts that reliable dialogue needs Does preference optimization harm conversational understanding?. Next-turn reward signals push the same way, teaching models to respond passively instead of actively probing for what the user actually means Why do language models respond passively instead of asking clarifying questions?. The downstream cost is measurable: across 200,000+ conversations, models lock into premature assumptions in underspecified, gradually-revealed dialogue and lose ~39% of their performance Why do language models fail in gradually revealed conversations?.
What's quietly striking is that this is a missing-mechanism problem, not just a missing-skill one. Collaborative Rational Speech Acts offers a formal model of the thing humans do effortlessly — bidirectional belief tracking that follows a conversation from partial to shared understanding — and demonstrates it on doctor-patient dialogue, precisely the information-theoretic scaffolding that token-by-token LLMs lack Can dialogue systems track both speakers' beliefs across turns?. And the gap isn't only interpersonal: models can articulate a correct principle (87% accuracy) yet fail to apply it (64%), a split between knowing and doing that mirrors, inside one model, the same disconnect between stating an explanation and successfully delivering one Can language models understand without actually executing correctly?.
The thread to pull, if you want to go further: structuring reasoning as dialogue rather than monologue measurably improves diversity and coherence Can dialogue format help models reason more diversely? — which hints that the interactive, multi-move structure the explanation study found in human conversation might be a feature worth engineering into models, not just a human quirk they're failing to imitate. The surprise underneath this whole line of work is that a good explanation may have less to do with being right and more to do with the back-and-forth that confirms you've been understood.
Sources 9 notes
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.
Explanation quality is not intrinsic to the explanation itself but depends on the rhetorical situation: who presents it, how it is framed, and what role the recipient plays. Evaluations that ignore this triad measure only a narrow slice of real-world effectiveness.
LLMs produce grounding acts—clarifications, acknowledgments, repairs—77.5% less frequently than humans. They generate fluent responses without verifying shared understanding, relying instead on authoritative framing that masks the absence of genuine communicative calibration.
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.
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.
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.
Large language models can articulate correct principles but systematically fail to apply them due to dissociated instruction and execution pathways. The 87% accuracy in explanations versus 64% in actions reveals this is not knowledge deficit but structural disconnect.
DialogueReason, which structures a single model's internal reasoning as dialogue between distinct agents in separate scenes, overcomes monologue reasoning's fixed-strategy and fragmented-attention weaknesses, especially on tasks requiring multiple problem-solving approaches.