Could proactive dialogue make conversations dramatically more efficient?

Proactivity in dialogue — providing relevant information even when not explicitly requested — is "very common in human-human dialogues" but "almost absent from current research in task-oriented dialogue systems." The data confirms this: proactivity is "largely under represented in most of the datasets" used to train and evaluate dialogue systems.

The example is simple but revealing:

• User: "What time is the next train to London?"
• Agent: "The next train is at 10:15. It arrives at 12:45."

The arrival time was not asked for, but the agent guesses (correctly) that this is information the user will likely need. This follows Grice's cooperative maxims — specifically, being informative enough to serve the conversational purpose.

Simulation experiments investigating four aspects of proactivity — degree of system proactivity, user influenceability, domain complexity, and user-need/domain fit — demonstrate that proactivity can reduce dialogue turns by up to 60% in medium-complexity application domains. This is not a marginal improvement; it fundamentally changes the efficiency of the interaction.

The absence from research is particularly striking given the efficiency gains. Since Why can't conversational AI agents take the initiative?, the passivity is not just a capability gap — it is a data gap. Models trained on datasets that lack proactive examples cannot develop proactive behavior even if the architecture supports it. The training signal simply isn't there.

This connects to a broader pattern: since Does preference optimization harm conversational understanding?, RLHF training specifically penalizes proactive responses (adding information the user didn't ask for can seem presumptuous to raters evaluating single turns), even though proactivity massively improves multi-turn efficiency.