Social robots that visibly wonder out loud make the people around them more curious. Users ask more questions, explore more, and seek out uncertainty. In every study where this effect was shown, though, the robot's curiosity was scripted: it voiced wonder at moments its designers chose, not at moments where the system itself was actually unsure. This proposal makes the agent's curiosity honest. A small committee of differently trained reading models (an ensemble) looks at the document and the user's question. When the members point to different passages, the agent asks a question that names the competing readings, for example "Do you mean the renewal deadline or the late-fee rule?" The agent stays quiet when the members agree.
The claim has two parts. The first is that disagreement between members pre-trained in different ways flags the turns where the best single model is about to ground its answer in the wrong passage, and flags them better than that model's own confidence does. The second is that a question built from the specific disagreement does two things that generic scripted wonder does not. It helps the user finish the task, and it still produces the curiosity contagion that earlier robot studies found. We test the first part offline on Doc2Dial and the second in a controlled user study. All uncertainty sources plug into one fixed interface, so every comparison is between interchangeable parts under the same question budget.
Background and the gap
Two lines of work on verbal curiosity motivate the user-facing side. Expression of Curiosity in Social Robots compared three NAO robot conditions in a rock-classification game. Participants with a robot that asked curious questions asked 9.67 on-task questions. Participants with a neutral robot asked 1.30. When the robot also explained why it was curious, participants asked only 3.40. Can Children Catch Curiosity from a Social Robot? found that children adopted only the specific exploratory behaviours the robot modelled. Free exploration and uncertainty seeking increased, but question generation did not. Children paired with the non-curious robot also showed larger immediate learning gains. Two lessons follow for our design. Contagion depends on which behaviour is modelled. And curiosity can cost task performance, which our design has to measure rather than assume away. In both studies the curiosity was scripted and had no link to the agent's knowledge.
Computational curiosity in dialogue exists, but it stays internal to the system. Stochastic Curiosity Exploration for Dialogue Systems rewards a dialogue policy for prediction error in a learned latent state. CURIO rewards a policy for reducing a single auxiliary language model's uncertainty about hidden user attributes. On exercise recommendation it reaches an 87.5% success rate, against 68.5% for standard multi-turn RLHF. Both papers treat curiosity as a training signal. Neither asks whether voicing it changes the human, and CURIO relies on one user model. CERMIC packages curiosity as a plug-in module for multi-agent reinforcement learning, but it is evaluated only in simulated control environments.
The ensemble side supplies our method. The DialDoc21 report from team SCIR-DT combined two differently trained models. One was a RoBERTa model further pre-trained on heavily augmented dialogue data. The other was an ELECTRA model deliberately left without that extra pre-training, so that it kept broad representations. Their weighted combination beat each member by over 4% in F1 and Exact Match. That gap means the two members make different errors, and our method uses exactly those differences as the trigger for a question. Ha et al. found that averaging diverse user models was the most consistently reliable predictor across tasks. They also found that every model degraded on open-ended exploration. That limits what we promise: disagreement should be informative on directed, document-grounded tasks, and we make no claim for open-ended discovery. SPIO showed that keeping several top candidates rather than selecting one improves robustness in automated data science. In every one of these papers, disagreement is averaged away to improve a score. None turns it into something the agent says.
The modularity side supplies discipline rather than novelty. OmniEvaluator showed that evaluation scores can swing by up to 80 points when only prompt formatting or the evaluation engine changes. Its fix was to fix the stage boundaries and use an independent verifier. UCEE gave causal estimators a shared fit/predict interface so they could be compared without custom glue code. DeepDIVA tied every run to a code commit and fixed random seeds. We adopt the same pattern for one component: the decision of when and what to ask.
Core claim
Let $p_k(s)$ be the probability that ensemble member $k$ (of $K$) assigns to candidate grounding span $s$, and let $\bar p$ be their average. The trigger score is the ensemble's disagreement. We measure it as mutual information, which is the entropy of the averaged prediction minus the average entropy of the individual predictions:
$$D = H(\bar p) - \frac{1}{K}\sum_{k=1}^{K} H(p_k)$$
Here $H$ is entropy over candidate spans. $D$ is large when each member is confident but the members favour different spans. A single model's own uncertainty cannot express this pattern. When $D$ exceeds a threshold, the agent asks a contrastive question built from the top span of each disagreeing member. The threshold is set so that every trigger source asks the same number of questions.
The claim is falsifiable on two fronts. Offline, $D$ should predict grounding errors of the best single model better than that model's own entropy or top-two margin. In people, contrastive questions triggered by disagreement should raise task success above a neutral agent. They should also produce at least as many user questions as scripted wonder asked at the same rate.
Research program
We build a document-dialogue agent from four stages with fixed interfaces. A grounding stage returns span distributions. A trigger stage receives those distributions and returns a decision to ask or not ask, together with candidate readings. A verbalizer turns the decision into an utterance. A logging and evaluation stage records everything against a commit hash and seed, following DeepDIVA.
The trigger interface is where the scientific comparison happens. Five trigger sources sit behind the same signature and the same question budget: ensemble mutual information, single-model entropy, single-model top-two margin, a language model's verbalized confidence, and random triggering. The verbalizer has three modes. It can produce generic wonder ("Hmm, I wonder about that part"), a contrastive question naming the two readings, or a contrastive question plus a one-clause rationale ("my two readers disagree here").
The ensemble starts as a reproduction of the SCIR-DT pairing, with RoBERTa adapted on dialogue data and ELECTRA left general. We then add a third member trained on a different augmentation view to test whether diversity, rather than member count, drives the signal. Building the ensemble this way lets us vary diversity directly. A same-architecture ensemble that differs only in random seed serves as the ablation.
Baselines
The naive reference point is an agent that never asks and simply answers with the combined ensemble's span. On Doc2Dial user-utterance grounding, the original dataset paper doc2dial reports an Exact Match of 55.1%. The DialDoc 2021 Shared Task reports a best Subtask 1 system at EM 67.1 and F1 76.3. That shared-task figure is the strongest reported grounding number we can cite, and any agent that asks questions must improve on its own no-question version, not just on this number.
The strongest competitor for the trigger is the best single grounding model's own top-two margin. It is computed for free and already yields two candidate readings to contrast. It runs on the same models, the same dev-set threshold-tuning budget, and the same matched question rate. On Lacuna we found no reported result for any question-asking trigger on Doc2Dial, so that comparison has to be produced here.
The baseline for contagion is scripted curiosity at a matched rate, the design of Ceha et al. Their reported figures are 9.67 on-task questions for the curious robot and 1.30 for the neutral one. Their setting differs from ours, so we treat these as a reference for effect size, not a target.
Two ablations isolate the key ingredient. The first uses a seed-only ensemble to test whether pre-training diversity matters. The second keeps the trigger fixed and swaps contrastive for generic wording, to test whether the content of the question matters. To count as beating the margin baseline, ensemble disagreement must achieve higher error-detection AUROC with a confidence interval that excludes zero difference. It must also achieve higher post-clarification EM at every matched question rate between 5% and 30% of turns.
Why not just…?
Why not just use the single model's own confidence? This is the real threat. A single model's top-two spans already provide a contrastive question, and its entropy is a reasonable trigger. The only evidence favouring the ensemble is indirect. The 4%-plus gain in the SCIR-DT report shows that the members make different errors, but that does not prove disagreement localizes errors better than one model's confidence. The offline experiment settles it directly. If the margin baseline matches ensemble disagreement on error detection and post-clarification EM, we drop the ensemble from the user study. The contribution then narrows to the second part of the claim, contrastive versus generic curiosity, which stands independently.
Why not just prompt a large language model to say when it is unsure? Verbalized confidence is cheap and needs no extra models. It is also sensitive to the prompt, the kind of fragility OmniEvaluator documented in evaluation. In our design it sits behind the same trigger interface and question budget as a fifth source, so the offline curves either show it competing or show it losing. It is not dismissed in advance.
Why not just insert scripted curious phrases? Prior studies show that scripted curiosity is contagious, so it is a serious baseline for the human outcome. It cannot help the task, though, because its questions carry no information about the document. The non-adaptive robot in the Mole Madness acoustic study also failed to show engagement-moderated synchrony with children. The user study includes scripted wonder at the same rate. If it matches disagreement-triggered questions on both contagion and task success, our claim fails.
Why not just train one end-to-end model that decides when to ask? Doc2Dial contains no labels for when a curious question should be asked. An end-to-end model would therefore need either reinforcement learning against a simulated user, as in CURIO, or invented supervision. Either route ties the trigger, the wording, and the grounding together, so we could no longer tell which part produced an effect. The modular design exists precisely to keep those effects separable. An end-to-end model is a reasonable follow-up once we know which ingredient matters.
Proposed experiments
The offline study uses the Doc2Dial user-turn grounding split. It reports AUROC for predicting whether the best single model's span is wrong, separately for each trigger source. It then runs a clarification simulation. When the agent asks, an oracle user picks whichever candidate overlaps the gold span, or replies "neither". We plot EM after clarification against the question rate for every trigger. All thresholds are tuned on dev data and all reported runs use five seeds. The decisive check is whether the ensemble curve lies above the top-two-margin curve.
The user study is a between-subjects online experiment. Participants complete document-grounded information tasks drawn from Doc2Dial's government-service domains while chatting with the agent. There are five conditions: neutral; scripted wonder at a matched rate; disagreement-triggered generic wonder; disagreement-triggered contrastive questions; and contrastive questions with a disagreement rationale. Sample sizes come from a power analysis based on the condition differences Ceha et al. observed. We aim well above their N=30 total.
The outcomes are the number of on-task questions users ask, task correctness, self-reported curiosity, and recall at one week. Because Engel et al. found that curiosity can cost immediate learning, we treat that trade-off as a primary outcome, not a nuisance variable.
The rationale condition addresses the tension in Ceha et al. directly. Their robot's explanations reduced question-asking. A one-clause statement of disagreement is shorter and more concrete than their explanations, so it may avoid that effect or reproduce it. Either result tells future agents whether to reveal their uncertainty or only act on it.
References
- Expression of Curiosity in Social Robots: Design, Perception, and Effects on Behaviour
- Can Children Catch Curiosity from a Social Robot?
- Stochastic Curiosity Exploration for Dialogue Systems
- Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward
- Wonder Wins Ways: Curiosity-Driven Exploration through Multi-Agent Contextual Calibration
- Technical Report on Shared Task in DialDoc21
- A Unified Comparison of User Modeling Techniques for Predicting Data Interaction and Detecting Exploration Bias
- SPIO: Ensemble and Selective Strategies via LLM-Based Multi-Agent Planning in Automated Data Science
- A Composable Evaluation System for Reproducible Omni-Modal Foundation Model Evaluation
- Universal Causal Evaluation Engine: An API for empirically evaluating causal inference models
- DeepDIVA: A Highly-Functional Python Framework for Reproducible Experiments
- doc2dial: A Goal-Oriented Document-Grounded Dialogue Dataset
- DialDoc 2021 Shared Task: Goal-Oriented Document-grounded Dialogue Modeling
- An Acoustic Analysis of Child-Child and Child-Robot Interactions for Understanding Engagement during Speech-Controlled Computer Games