Learning to Clarify Underspecified Intents Under Limited Interaction

1 Introduction

• Question value depends on the user’s goal: details about a table may matter more than breakfast food when the image is intended for a furniture catalog.Clarification asks about unarticulated aspects of a desired result so they can be fulfilled.Source: §1 · p. 2
• The paper formulates clarification as a value-of-information problem: a question is valuable to the extent that eliciting information improves the expected outcome for the user.The value is the difference between the best expected outcome after asking and the best expected outcome using information already available.Source: §1 · p. 2
• Training starts from a complete intent, decomposes it into atomic details, redacts some details, and rewrites the remainder as a naturalistic request.A simulated user retains the full intent and reveals withheld information in response to questions; user patience can limit the interaction budget.Source: §1 · pp. 2–3
• The framework is instantiated in text-to-image generation, where hyper-detailed descriptions provide complete specifications and outputs can be rendered for evaluation.Qwen3-8B is post-trained with simulated users of varying patience and evaluated under varying question budgets.Source: §1 · pp. 2–3
• The authors report improved output fidelity in simulated and real-user experiments, with gains attributed to better interaction rather than better guessing and, under some metrics, improved efficiency.The pipeline uses task descriptions rather than additional human interaction data for training.Source: §1 · p. 3

2.1 The value of asking

• Value-of-information work evaluates questions by the improvement their answers facilitate, including through sequential preference elicitation and downstream task outcomes.This paper extends that line by explicitly using omitted information for training and evaluation, varying underspecification and interaction budgets, and testing transfer to real users.Source: §2.1 · pp. 3–4

2.2 Proactive clarification

• Earlier proactive-clarification systems use heuristics or prompting to decide whether and what to ask; this work derives a measurable objective for training those decisions.The setup enables systematic comparison across question budgets, underspecification levels, and potential user behaviors.Source: §2.2 · p. 4

2.3 Preference feedback

• Unlike preference-feedback approaches that assume essential image elements appear in the initial prompt, this work allows important details to be absent and trains assistants to elicit them.The authors describe combining the approaches to scaffold multiple levels of outcome specification as a speculative possibility for future work.Source: §2.3 · pp. 4–5

2.4 Learning via simulated interaction

• Simulated interaction makes reinforcement-learning experiments feasible without requiring people in every training interaction, and the paper extends this approach to open-ended generation with withheld specifications.Evaluation with users measures output quality and interaction burden; the authors contrast this with prior work’s five-participant preliminary study and simplified protocol.Source: §2.4 · p. 5

3 Methods

• The framework trains an assistant on recovered utility from simulated-user interaction and evaluates it with real users.Task instances begin from complete specifications, with omitted information hidden from the assistant.Source: §3 · p. 5

3.1 Problem setup

• Complete specifications are decomposed into weighted atomic details, then some details are redacted to create requests with measurable missing utility.The assistant sees only the request and must query a privileged simulator that has access to the complete specification.Source: §3.1 · p. 6
• Three asking modes separate inference without questions, a fixed one-question-per-hidden-detail budget, and open asking with a stop decision.Open asking allows up to 2k questions, while the assistant does not know the hidden-detail count or budget in advance.Source: §3.1 · p. 7

3.2 Training to maximize the recovered utility

• The reinforcement-learning objective maximizes expected recovered utility, using group-centered rewards to favor recovery beyond what the request alone suggests.Patient- and impatient-user training differ in whether the assistant is rewarded after k answers or within an unknown stopping limit.Source: §3.2 · p. 8

3.3 User study

• The preregistered human study recruits screened Prolific participants for six randomized task rounds, with users answering until they stop or reach 15 questions.Participants try to reconstruct a viewed reference image; the image is generated once after the interaction.Source: §3.3 · pp. 8–9

3.4 Evaluation and analysis

• Evaluation combines simulated-user tests, which estimate recovered utility from known redactions, with human interactions that assess real-user performance.The simulated evaluation supports controlled comparisons across assistants, asking policies, and user patience levels.Source: §3.4 · p. 9, §3.4.2 · p. 10

3.4.1 Evaluation conditions.

• Simulated tests use 78 held-out DOCCI instances across base, patient-trained, and impatient-trained assistants, with fixed- and open-asking regimes.Both patient and impatient simulators are tested; no-asking conditions provide a baseline, while human comparisons include only the base and patient-trained assistants.Source: §3.4.1 · pp. 9–10

3.4.2 Outcome measures.

• Outcomes include hidden-detail recovery, semantic similarity changes across text and image comparisons, and interaction counts, effort, duration, and token use.Regression estimates use reference-image fixed effects and cluster-robust standard errors, with question-count adjustment omitted for count outcomes and simulated overviews.Source: §3.4.2 · p. 10, §3.4.3 · p. 10

4.1.1 Can models learn to ask more effective questions?

• 37.6–57.7 percentage-point adjusted recovered-utility gains over base appeared across all eight simulated asking comparisons.All comparisons had pBH < .0001.Source: Figure 2 · p. 11, §4.1.1 · p. 11

4.1.2 Do better questions lead to better output fidelity?

• Training improved reference-similarity gains across all four simulated text/image comparisons, with advantages ranging 0.040–0.195 for patient-trained and 0.016–0.113 for impatient-trained models.All comparisons had pBH < .05.Source: §4.1.2 · pp. 11–12, Figure 3 · p. 12

4.1.3 Can models decide how much information to elicit?

• Open asking reduced answered-question counts by 0.45–4.78 and lowered observed recovery by 7.2–13.3 percentage points across settings.Three of four count reductions were significant; the exception was patient-trained under impatient users.Source: §4.1.3 · pp. 12–13, Figure 4 · p. 13

4.1.4 How does user behavior affect performance?

• Impatient-trained models showed no clear advantage with impatient evaluation users; recovery was lower than for patient-trained models under both fixed and open asking.Observed recovery was 65.9% versus 69.1% under fixed asking and 57.7% versus 61.9% under open asking.Source: §4.1.4 · p. 13

4.1.5 How might this impact interaction burden?

• Under fixed asking, training lengthened questions by 21.9–36.8 words and answers by 7.2–10.3; open-asking advantages remained 6.0–17.2 points after interaction-length controls.Word counts are an imperfect proxy for user effort.Source: §4.1.5 · pp. 13–14, Figure 5 · p. 14

4.2.1 Do simulated-trained policies improve results from interaction with people?

• 0.0141–0.0337 cosine-similarity advantages over base appeared across human text/image comparisons after adjustment for question count and source image.All contrasts had pBH < .005.Source: Figure 6 · p. 15, §4.2 · p. 15

4.2.2 Do gains persist under comparable amounts of interaction?

• At 3–5 answers, trained assistants had significant advantages across all four human quality measures; at four answers, text–text advantage was 0.057 and image–image advantage was 0.023.No comparison was significant at zero or one answer.Source: Figure 7 · p. 16, §4.2.2 · p. 16

4.2.3 How does this affect the realized burden on actual users?

• Participants answered 2.27 fewer questions and rounds were faster, despite 72.3 more observed dialogue words; total token use fell by 1,420 per round.The interaction measures do not establish reduced subjective cognitive load; reasoning tokens decreased while visible-output tokens increased.Source: §4.2.2 · p. 16, §4.2.3 · p. 16, Figure 8 · p. 17

5.1 Eliciting authorship through participation

• Clarification can return otherwise automated decisions to users, potentially making their expressive intent more operative in generated artifacts [27].The authors connect this to authorship through participation.Source: §5.1 · p. 18

5.2 Without clarification, missing intent may leave room for undesirable influence

• Unresolved preferences may leave assistant decisions more open to undesirable priors or external nudges [14].The authors suggest investigating whether clarification constrains sensitivity to such influences.Source: §5.2 · p. 19

5.3 Learning to establish common ground

• Users may omit details even when they can inspect a reference throughout, so clarification can help them decide what to communicate.The authors argue this can reduce the need to anticipate every way an assistant might misunderstand an initial request.Source: §5.3 · p. 19

5.4 Mutually improving simulation and human-AI interaction

• Simulated users can teach useful interaction strategies despite imperfectly reproducing human behavior; observed stopping patterns motivate modeling patience around the value of further information.The authors hypothesize that users may stop once they have communicated enough useful information.Source: §5.4 · pp. 19–20

• Utility is operationalized as recovery of omitted details, while embedding similarity cannot establish their relative value or greater subjective satisfaction.The authors call for operationalizations sensitive to different users' goals and circumstances.Source: §6 · p. 20
• The reference-reconstruction task uses a stable, inspectable target, so extension to evolving preferences or decisions users want assistants to make requires further experimentation.The authors also identify tolerable variation as a setting not covered by the task.Source: §6 · p. 20
• Transfer is bounded by evaluation in one task domain with one assistant model against its untrained counterpart.The authors call for testing other simulators, generators, and clarification methods.Source: §6 · p. 20

7 Conclusion

• Task specification can remain challenging as task execution becomes more automated; the authors envision assistants that make users’ knowledge and judgment more consequential.They invite the community to build on their framework toward that ambition.Source: §7 · p. 20

A Examples

• Figure 10 shows examples from trained and untrained models beside the reference and request images, while Figure 11 presents a complete trained-assistant episode.The episode runs from the request through the judge’s verdicts.Source: A Examples · p. 24

Base Model

• Figures 10 and 11 illustrate generated images for six test instances and one patient-trained assistant episode, respectively.The examples place outputs beside a reference photograph and an image generated from the request alone; the episode runs from the reference image to the judge’s verdicts.Source: Figure 10 · p. 24, Figure 11 · p. 25

B Instance Synthesis

• The dataset starts from the first 2,000 DOCCI descriptions of at least 400 characters, each paired with its described photograph as the reference image.Each instance is built from one description, which passes through decomposition and redaction models; the same source always yields the same instance.Source: B Instance Synthesis · p. 26
• Decomposition leaves 479 instances with 20–45 details each after excluding 1,521 sources that yielded fewer than 20 details.Each detail has a self-contained sentence, an importance weight, and a binary judge check; at least five details remain after hiding up to 15.Source: B Instance Synthesis · p. 26
• Hidden details are sampled without replacement, with more important details more likely to be hidden, but every detail remains reachable.Across 479 instances, 3,791 details were hidden, averaging 7.9 per instance and distributed across hidden counts of 1 to 15.Source: B Instance Synthesis · p. 26
• The redaction model rewrites the kept details as a one-to-three-sentence user request that omits every hidden detail.In the example instance, the request states 14 kept details and omits 8 hidden ones; Table 1 presents a request and hidden details for one test instance.Source: B Instance Synthesis · p. 26, Table 1 · p. 26
• Sources are assigned by a seeded hash to training, validation, and test splits of 290, 111, and 78 instances, respectively.The split shares are 0.6, 0.2, and 0.2.Source: B Instance Synthesis · p. 26

C Training and Evaluation Details

• The assistant is Qwen3-8B, while the decomposition, redaction, user simulator, and judge each use a structured Gemini 3.5 Flash-Lite call.Assistant calls sample at most 4,096 tokens, shared between reasoning and visible text; failed environment calls are retried up to three times.Source: C Training and Evaluation Details · p. 27
• The judge checks whether the composed description states or clearly entails each hidden detail, and recovery is computed under importance and uniform weights.The two weighting schemes move closely together in the reported evaluations.Source: C Training and Evaluation Details · p. 27
• Importance- and uniform-weighted recovery scores move closely together and produce identical rankings across the 15 test evaluations.The mean importance-weighted score exceeds the uniform-weighted mean by 0.007 on average and at most 0.014; the authors report no change in comparisons between assistants, modes, or users.Source: C Training and Evaluation Details · p. 27
• Training uses 8 instances per iteration and 32 episodes per instance, with each rollout’s recovery reward centered against its instance group mean.For impatient-user training, all 32 rollouts for an instance and iteration face the same sampled patience; the group baseline accounts for instance difficulty and lowers gradient variance.Source: C Training and Evaluation Details · p. 27
• The update descends an importance-sampled policy-gradient loss over sampled episode tokens, including questions and descriptions but excluding prompt tokens.The policy is updated once per batch, so the sampling-to-training weight ratio is one; the loss uses REINFORCE with a group baseline and no PPO clipping, KL penalty, entropy bonus, or length normalization.Source: C Training and Evaluation Details · p. 28
• Checkpoints are saved every five iterations, and validation selection evaluates them in fixed-asking mode against the training user.The patient run selected iteration 219 on the validation plateau; the impatient run selected iteration 294 as its best score under the impatient user.Source: C Training and Evaluation Details · p. 28
• Evaluations pair impatient-user patience across assistants and modes, while generated images and text-image similarities use specified rendering and embedding settings.Images are rendered at 512 × 512 pixels; similarities use cosine similarity between Gemini Embedding 2 embeddings.Source: C Training and Evaluation Details · pp. 28–29

D User Simulator Audit

• The audit covers 667 answers per assistant from the 78 test instances, with a judge classifying whether each answer states a hidden detail and whether the question asked for it specifically, by topic, or not at all.The simulator is instructed to answer only what was asked, one detail at a time, deflect open-ended questions, and state no preference when the description is silent.Source: D User Simulator Audit · pp. 29–30
• In the audited patient-user, fixed-asking conditions, nearly every hidden detail stated by an answer was asked for specifically or by topic, and volunteered details were rare.The authors conclude that trained assistants largely recover details by asking rather than exploiting the simulator; Table 3 summarizes the audit verdicts.Source: Table 3 · p. 29, D User Simulator Audit · p. 30

E User Study Materials

• The study recruited participants through Prolific and collected their typed text, assistant questions, generated images, timestamps, screening answers, and Prolific IDs.The consent materials describe an approximately 45-minute study and a bonus based on how closely generated images match their targets.Source: E User Study Materials · p. 30
• Participants saw a target photo, wrote a request, answered questions one at a time, and could stop and generate after any question, up to 15 questions.The image was generated once at the end; participants were instructed to match the target as closely as possible.Source: E User Study Materials · pp. 30–31, Figure 13 · p. 31

F Auto-Interpretability Results

• The auto-interpretability analysis recursively clusters descriptions of how trained-model questions and participant answers differ from the base model.It reports second-from-top-level themes in Tables 4 and 5, with example quotations.Source: F Auto-Interpretability Results · p. 31, Table 4 · p. 32, Table 5 · p. 32

G Regression Tables

• The regression tables report simulated and human-study contrasts, with coefficient p-values unadjusted unlike the BH-adjusted figure contrasts.The appendix says all regression models follow the estimation approach of Section 3.4.Source: G Regression Tables · p. 32

G.1 Model Results

• Tables 6–9 cover simulated recovered utility, similarity gains, and dialogue measures, using source-image fixed effects and clustered standard errors.The base model with fixed asking and the patient user is the reference level.Source: §G.1 · p. 32, Table 6 · p. 33, Table 7 · p. 34
• The simulated regressions add interactions among assistant, asking mode, and user in successive specifications, then add dialogue measures as covariates.The appendix notes that coefficient p-values are unadjusted, unlike the BH-adjusted figure contrasts.Source: §G.1 · p. 32
• Tables 10–14 cover human-study similarity gains and process measures, with participant-clustered standard errors.The similarity regressions are reported both without and with adjustment for questions answered and source-image fixed effects.Source: §G.1 · p. 37, Table 10 · p. 37, Table 11 · p. 37

H Additional Results

• Under open asking, trained assistants ask in most episodes but stop before the fixed-asking budget; recovery is also broken down by hidden-detail count.Figures 14 and 15 show recovery by hidden details and the share of episodes reaching each answered turn, respectively.Source: H Additional Results · pp. 38–39, Figure 14 · p. 39, Figure 15 · p. 39

I.1 Assistant

• The assistant prompts ask one short question about what matters most and remains unclear, decide whether to continue, and compose a complete image description.The description prompt requires retaining stated details and filling visually important unspecified details with likely choices.Source: §I.1 · p. 40

I.2 User Simulator

• The simulated user answers truthfully from the complete intent, avoids volunteering details, and may stop answering according to its behavior setting.For open-ended questions it reveals no detail, and for unspecified intent details it says it has no preference.Source: §I.2 · pp. 40–41

I.3 Judge

• The judge marks a check true only when the image description explicitly states or clearly entails it, without relying on outside knowledge.Silence, contradiction, or vague compatibility counts as false.Source: §I.3 · p. 41

I.4 Image Model

• The image model prompt requests a single image rather than multiple options.Source: §I.4 · p. 41

I.5 Synthesis

• The synthesis prompts decompose complete descriptions into atomic, checkable details, each with a category, importance score, and verification check.They also redact selected details to produce a natural request while preserving all remaining facts and avoiding hints about removed details.Source: §I.5 · p. 42
• The redacted-request prompt specifies a casual, compact request of one to three sentences that excludes the removed details.The output is the natural-language request stating everything in the description except the redacted details.Source: §I.5 · pp. 42–43