Why do monological explanations fail to transfer understanding compared to dialogical ones?
This explores why explanations delivered as one-way monologue transfer understanding less reliably than explanations built back-and-forth in conversation — and what the corpus says about where the gap comes from.
This explores why explanations delivered as a finished monologue transfer understanding less reliably than ones co-built in dialogue. The corpus points to a single root cause: understanding isn't a payload that gets shipped intact from speaker to listener — it's something assembled jointly, in real time, by both parties. The clearest evidence comes from analysis of nearly 400 everyday explanations What makes explanations work in real conversation?, which finds that whether an explanation lands depends on three things interacting — how it relates to the current topic, the dialogue act it performs, and the explanatory move it makes. A monologue fixes all three in advance, before it knows where the listener actually is. Dialogue lets them be renegotiated turn by turn. That's the mechanism: the monologue optimizes for a generic listener who doesn't exist.
A second framing reframes the whole problem as communication rather than content. The work on explainable AI What if XAI is fundamentally a communication problem? argues that explanation quality is never intrinsic to the explanation itself — it lives in the triad of who delivers it, how it's framed, and what role the recipient is given. A monologue strips out two of those three legs: the recipient becomes passive, and the framing can't adapt. So an explanation that is objectively complete can still fail, because completeness was never the variable that mattered.
Here's the part you might not expect: this isn't only a human-listener problem — it's baked into how today's models are trained to talk. Preference optimization (RLHF) rewards confident, single-turn answers and penalizes the very moves dialogue runs on — clarifying questions, checking whether the listener followed. One study finds these "grounding acts" drop by over 77% below human levels Does preference optimization harm conversational understanding?. We've effectively trained models to prefer monologue, which makes them appear helpful while failing silently the moment understanding needs to be verified rather than asserted.
Laterally, the corpus suggests monological transfer fails even when the speaker genuinely knows the material — because articulation and application live in different places. Models can state a concept correctly and then fail to use it, a "potemkin" pattern where explanation and execution are functionally disconnected Can LLMs understand concepts they cannot apply?, echoed by the comprehension-without-competence split Can language models understand without actually executing correctly?. A monologue only ever exposes the articulation pathway. Dialogue forces the explainer back into application — answering a follow-up means re-deriving, not reciting — which is also why reconstructing the hidden reasoning behind expert text, rather than just the polished text itself, transfers reasoning better Can reconstructing expert thinking improve reasoning transfer?. The finished explanation is a surface residue; the understanding was in the process that produced it.
The quietly surprising takeaway: dialogue may transfer understanding better not because it adds information, but because it repeatedly forces both parties to surface and repair the gaps a monologue is structurally blind to. The conversation isn't the delivery vehicle for the explanation — it's where the explanation actually gets made.
Sources 6 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.
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.
Models can explain concepts accurately, fail to apply them, and recognize the failure—a triple pattern incompatible with human cognition. This indicates functionally disconnected explanation and execution pathways rather than simple knowledge gaps.
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.
Training on expert texts augmented with reconstructed thought processes (self-talk, knowledge recall, verification) produces reasoning skills that transfer across domains and adapt depth to problem difficulty, outperforming standard continual pretraining by up to 8 points on hard problems.