How might dual-process dialogue use information gain to trigger clarification?
This explores how a dialogue system that switches between fast intuition (System 1) and slow strategic planning (System 2) could use a measure of 'how much would I learn by asking?' to decide when to stop and request clarification.
This explores how a dialogue system that switches between fast intuition (System 1) and slow deliberate planning (System 2) might use information gain — a measure of how much an answer would reduce its uncertainty — as the trigger for asking a clarifying question. The corpus doesn't contain a single paper that wires these three pieces together, but it holds all three pieces, and laying them side by side suggests a real design.
Start with the switch itself. Dual-process dialogue planning runs a cheap neural policy for familiar contexts and escalates to MCTS planning only when the model's own uncertainty estimate spikes Can dialogue planning balance fast responses with strategic depth?. That uncertainty signal is already doing the work of deciding 'is this routine or novel?' — the natural extension is to make the same signal decide 'should I answer or ask?' High predicted information gain from a clarifying question is essentially the formal version of 'I'm too uncertain to commit, and the user holds the missing bit.'
The missing bit needs a unit, and that's where the information-theoretic dialogue work comes in. Collaborative rational speech acts fuse rate-distortion theory with pragmatic reasoning to track both speakers' beliefs as they move from partial to shared understanding Can dialogue systems track both speakers' beliefs across turns?. That bidirectional belief model is exactly what you'd need to *estimate* expected information gain before asking — you can only score a question's value if you model what the other person knows and what you'd update. Pair that with the finding that clarifying value should be measured over the whole interaction, not the next token: next-turn reward optimization trains models to respond passively because immediate helpfulness is rewarded and the long-horizon payoff of asking is invisible Why do language models respond passively instead of asking clarifying questions?, and the same single-turn preference tuning erodes grounding behaviors by nearly 78% below human levels Does preference optimization harm conversational understanding?. Information gain is precisely the long-horizon quantity those reward schemes throw away.
There's a subtlety the corpus flags that an information-gain trigger has to respect: clarification doesn't have to be a question. Mapped onto Clark's ladder of communication, most real clarifications are declarative — a restatement, a hedge, a check — not a syntactic question, which makes them invisible to systems detecting clarification by sentence form Why do clarification requests look different at each communication level?. So a well-built trigger fires on the *information state* (uncertainty is high, the user can resolve it) rather than on a question template, and conversation-analytic 'insert expansions' give a formal vocabulary for when an agent should pause and consult the user instead of silently guessing When should AI agents ask users instead of just searching?.
The payoff worth knowing: clarifying isn't a tax on efficiency, it's a path to it. Proactivity — volunteering the right information, or extracting it — cuts dialogue turns by up to 60% in medium-complexity domains, yet is almost absent from AI benchmarks Could proactive dialogue make conversations dramatically more efficient?. An information-gain trigger inside a dual-process loop is one concrete way to earn that 60%: ask only when the expected reduction in uncertainty beats the cost of an extra turn, and stay in fast mode otherwise.
Sources 7 notes
A framework combining a neural policy model (System 1) for familiar contexts with MCTS planning (System 2) for novel scenarios, switching based on the model's own uncertainty estimates, matches or exceeds pure MCTS performance while reducing computational cost.
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.
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.
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 maps clarification mechanisms to four levels of communication—attention, signal, meaning, action—each grounded in a different modality (socioperception, hearing, vision, kinesthetics). Most clarifications use declarative form, not questions, making them invisible to systems that detect by syntax alone.
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.
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.