acceptodds
Under review as a conference paper at ICLR 2027

PAUSE: Proactive Appropriateness and Utility for Selective Elicitation

Abstract

Information-seeking agents must decide not only what information to acquire, but also whether a candidate intervention is appropriate for the current dialogue state. Existing clarification and value-based approaches primarily optimize expected information gain or task utility, implicitly assuming that a valuable intervention is suitable whenever it is selected. We show that this assumption can fail: the same question may be informative but premature at one point in a dialogue and appropriate at another. We introduce PAUSE, a selective elicitation framework that decomposes intervention selection into two decisions: contextual appropriateness and information utility. PAUSE learns an appropriateness score and a utility score from complementary supervision, and performs appropriateness-constrained utility optimization by selecting the highest-value intervention among candidates that satisfy a contextual fit constraint. On NewsInterview, PAUSE improves eight-turn information coverage from 0.5042 with utility-only selection to 0.5676, while leading to more selective clarification behavior on ClariQ. These results suggest that modeling contextual fit as a separate decision factor can improve proactive information seeking, particularly in extended multi-turn interactions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Proactive Agents, Information Seeking, Conversational Decision Making

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_grz3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper looks at interactive dialogue settings where LLMs must ask questions in order to seek information or resolve ambiguity. The authors argue that standard approaches based on expected information gain alone “may not capture whether a question is premature, redundant, or contextually misaligned.” Accordingly, they propose a dual-objective framework, PAUSE, that introduces a learned “appropriateness” score. On two multi-turn benchmarks (NewsInterview and ClariQ), they report PAUSE leads to modest gains in coverage and selectivity.

Critical strengths
  • The information-seeking setting is both general and relevant to many applications beyond the ones considered here; e.g., similar approaches have recently been applied to clinical diagnosis and scientific discovery.
  • The experiments include a nice set of ablations (different hyperparameters and thresholds) as well as prompt-only baselines and different model backbones.
  • The method is clearly described. The writing is relatively readable and free of obvious AI slop, though too much space is allocated to certain figures (e.g., Fig. 2) and tables.
Critical weaknesses

My main concern here is that the method is solving a problem that I am not convinced actually exists outside the authors’ specific setting. The central premise is that PAUSE addresses contextual prematureness; in the authors’ words:

“A question can have high prospective information value while presupposing an event or preference not yet established.” (L165)

The problem here is that LLMs are already extremely adept at exactly this kind of pragmatic inference. This is a direct consequence of the standard next-token prediction objective, which encourages models to learn dialogue structure in which utterances follow naturally without presupposing events or preferences.

The kind of contextual inappropriateness that PAUSE addresses arises primarily in artificial benchmark settings where there is a fixed candidate set R=R1,…,Rk\mathcal{R} = {R_1, \ldots, R_k} in which some of the answers contain presuppositions or are mutually-redundant.

When there is a learned generator π(R∣S)\pi(R \mid S), as is the case here, the presence of contextually-inappropriate responses suggests that something in the setup has caused the LLM to diverge from its base policy: either the generator prompt leaks privileged information about the setting, or the training procedure itself causes divergence. Either way, simply sampling candidate responses from a frozen LLM, and then reranking them by EIG, avoids these pitfalls and is already a well-established approach (e.g., 1-3 below).

In sum, I’m unconvinced by the need for a multi-objective optimization framework when simple sample-and-rerank methods already work quite well. In order to motivate this work, the authors would need to better establish that the problems they claim to solve occur in natural dialogue (not just benchmark settings) and as part of standard EIG-based methods (not just as a byproduct of utility optimization).

References

  1. Mazzaccara et al., 2024. Learning to Ask Informative Questions: Enhancing LLMs with Preference Optimization and Expected Information Gain. https://arxiv.org/pdf/2406.17453
  2. Choudhury et al., 2026. BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design. https://arxiv.org/abs/2508.21184
  3. Grand et al., 2026. Shoot First, Ask Questions Later? Building Rational Agents that Explore and Act Like People. https://arxiv.org/pdf/2510.20886
Questions for authors
  1. How robust is your method to different hyperparameter configurations? If you were to deploy this in production, how would you choose from the many reported combinations of coefficients and thresholds (Table 1)?
  2. In addition to presupposition, another problem you point to with value-of-information methods is redundancy. However, under a calibrated belief model, redundant questions should yield low information. Can you give some intuition for why this is not the case in your setting?
3 comments
AuthorsOpenReview ↗

Response to Reviewer grz3

Thank you for the connections to Bayesian information gathering and the questions about candidate generation, training, and redundancy. We added controls that distinguish the effects of candidate construction, utility estimation, and contextual selection.

Are the candidates or training procedures creating the problem?

NewsInterview samples four new questions from the current dialogue state at every turn. The generator and selector receive introductory context and dialogue history; the reference transcript and information-item list are reserved for the respondent and evaluator. ClariQ is a separate, single-turn candidate-selection task. The SPoRC position probe uses recorded human dialogue with controlled position perturbations to evaluate the scorer's placement sensitivity. We will clarify these distinct experimental roles and input boundaries.

Removing T-based filtering from generator training retains a gate gain of +0.0795 [0.0467, 0.1131]. The benefit therefore persists without using T to construct the training examples. This control retains utility-oriented SFT/DPO and specifically tests the role of the T filter.

A fully unadapted generator with a standard chat template and learned-utility reranking reaches 0.5055, compared with 0.5606 for the trained-and-gated system, an observed difference of +0.0551 from the reported means. This comparison changes both generation and selection. Within the frozen standard-chat variant, the gate adds +0.0038, which is not statistically significant. Thus, the measured value of the gate depends on the generator and prompt; our evidence supports the tested selection policies rather than a universal need for gating.

Why can utility scoring favor redundant questions?

Under a calibrated belief model, a question that adds no evidence about the latent variable should have low conditional information gain. The distinction in our setting is that V predicts an outcome-derived novelty target. The original scorer takes the question without an explicit representation of the information already established in the dialogue. This proxy can therefore assign a high score to a question whose answer is generally informative but already known in the current conversation.

The cited methods make this distinction important. BED-LLM constructs a probabilistic model and updates beliefs to estimate EIG; Grand et al. combine Bayesian information gathering with action selection. We will position PAUSE as contextual selection with learned utility proxies and distinguish it from these Bayesian formulations. Our frozen-generator control retains the learned scorer and therefore does not constitute a reproduction of posterior-based EIG or establish superiority over it.

Do stronger contextual baselines absorb the benefit?

The gate still improves the tested history-aware utility scorer by +0.0349 [0.0015, 0.0683] (paired 95% bootstrap CI). A complementary deduplication control improves the original utility-only baseline by +0.0517 [0.0166, 0.0869], reaching 0.5560 versus PAUSE's 0.5606; the remaining +0.0047 difference is not statistically significant. These controls establish a strong lexical deduplication baseline and a gate benefit in the history-aware configuration.

Soft score combinations provide another useful comparison. The rules 0.25T + 0.75V and 0.50T + 0.50V reach 0.5529 and 0.5478, respectively, above utility-only selection at 0.5042. Their differences from hard filtering are not established as statistically significant. We will present hard filtering as an interpretable eligibility constraint and soft weighting as a competitive alternative. The supported contribution is contextual selection with learned utility; timing-specific attribution remains a separate mechanism question.

How should the threshold and candidate budget be chosen?

The gain is positive across all six tested thresholds between 0.40 and 0.60, with four intervals excluding zero. The canonical NewsInterview threshold, τ = 0.47, was fixed before the main evaluation and remains unchanged; the sweep is a sensitivity analysis rather than a procedure for selecting a new test-optimal threshold.

At K = 2, 4, and 8, the gate effects range from approximately +0.030 to +0.056, with intervals excluding zero at K = 4 and 8. This supports the rule at multiple candidate budgets, although a production budget also requires latency measurements.

For deployment, we recommend fixing the scorer configuration and threshold on development data using an explicit operating target, then evaluating the frozen policy on test data. The weighted-score coefficients belong to comparison rules, not to PAUSE's hard-filter policy. We will make this distinction and the sensitivity results explicit in the experimental description.

I'm not sure if it's just me but I found this rebuttal a bit hard to parse. As a concrete example:

Removing T-based filtering from generator training retains a gate gain of +0.0795 [0.0467, 0.1131].

What is T-based filtering? What is a gate gain? Is the quoted value a positive or negative result?

After reading through the comment several times, I'm still not much clearer on most of the questions I asked. I think LLM writing may be making it harder than necessary for me to see how these comments address my questions or the questions of the other reviewers.

In the absence of a clear rebuttal, I'm going to retain my original score.

Thank you for your reply. We will explain the quoted result first, then answer your original questions directly.

What the quoted experiment means. T is the score from a separate model that predicts whether a question fits the current conversation. During training, we normally use it to exclude questions the model judges unsuitable.

We then tested that same generator on 348 simulated interviews, with eight questions per interview. Coverage is the average percentage of reference facts recovered; higher is better.

For the second rule, if no question passes the fit check, we choose the one with the highest fit score. By “gate gain” we meant this difference: 0.0795, or +7.95 percentage points, with a 95% confidence interval of +4.67 to +11.31 points. This is a positive result. It shows that checking fit when selecting questions helps even without fit-based filtering during training.

Why redundant questions can receive high scores. Under a calibrated belief model, redundant questions should have low EIG. Our original predictor scores a question without explicitly representing what the dialogue has already established. It can therefore overvalue information already known. A simple repetition filter reaches 55.60% coverage versus PAUSE’s 56.06%.

Whether the setting creates the problem. At each NewsInterview turn, four new questions are generated from the introduction and dialogue history, rather than taken from a fixed question bank. Neither the generator nor the selector sees the hidden reference facts.

How the threshold is chosen. PAUSE uses one fit-score cutoff, fixed at 0.47 before the main NewsInterview test. Across six tested cutoffs from 0.40 to 0.60, all coverage differences were positive and four were statistically significant. For deployment, we would choose this cutoff on development data. The weighted-score coefficients in the table belong to comparison methods, not to PAUSE’s selection rule.

Reviewer_Mr9ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PAUSE, a framework for selective elicitation that decomposes intervention selection into two decisions: context appropriateness and information utility. The framework learns scores for each decision with complementary supervision and selects the highest-value intervention among candidates that satisfy a contextual fit constraint. The authors also fine-tune an intervention generator via SFT and DPO to maximize utility. Finally, they evaluate the method on NewsInterview and ClariQ, finding it improves information coverage and more selective clarification, respectively. However, the optimal mechanism for combining appropriateness and utility remains open.

Critical strengths

The motivation is sound, the experimental setup is well delimited, and the statistical treatment is satisfactory (although Figure 2 could use some error bars).

Critical weaknesses
  • The gains are significant but not by a large margin, and it seems hard to justify the added complexity of the method against simple prompting. Also, the necessity and impact of training the generator is not empirically justified.
  • The evidence is thin and the methods support only setups where the labels for informativeness and appropriateness are known. While this is acknowledged by the authors and they don't overclaim over it, it is unclear whether the proposed mechanism would have potential for broader impact and work for other domains.
  • Some terms are used without proper introduction, like MRR and UAR.
Questions for authors
  • Can you demonstrate the impact that training the generator has on the experiments and the main claims?
  • Can you discuss the trade-off between the method's added complexity against prompting?
  • Why does the pipeline with Qwen3-14B perform worse than with Qwen3-8B as scorer?
  • How would this work when we don't have labels for informativeness and appropriateness, on open-ended settings, for example?
  • What do MRR and UAR mean?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Mr9Z

Thank you for recognizing the motivation, experimental controls, and statistical treatment. The additional experiments separate the contribution of inference-time selection from generator training and make the performance–deployment trade-off more explicit.

What does generator training contribute?

The main gain is measured by changing the selection rule while holding the generator fixed. On 348 NewsInterview conversations, Coverage@8 increases from 0.5042 to 0.5606, a paired difference of +0.0564

00˙216,00˙9120\.0216, 0\.0912

(95% bootstrap CI).

The inference-time benefit also persists when T is removed from training-data construction. With that generator, the gate adds **+0.0795

00˙467,01˙1310\.0467, 0\.1131

**. Under gated selection, generators trained with and without the T filter differ by +0.0029, which is not statistically significant. This control retains SFT/DPO and specifically tests the training filter, rather than removing all generator adaptation.

A frozen base generator with a standard chat template and learned-utility reranking reaches 0.5055, compared with 0.5606 for the trained-and-gated system, an observed difference of 5.51 percentage points from the reported means. Frozen and trained utility-only selection reach 0.5055 and 0.5042, respectively, a frozen-minus-trained difference of +0.0013. Thus, the evidence supports the combined system and the controlled selection-rule gain, while a standalone training advantage over the standard-chat baseline is not established. We will present generator adaptation as a separate design choice and organize the ablations accordingly.

Does the benefit justify the added computation?

PAUSE improves on the strongest direct prompt-only interviewer under the same conversation and question budget. GLM-5.1 reaches 0.5227, compared with PAUSE's 0.5606, a paired difference of **+0.0379

00˙067,00˙6920\.0067, 0\.0692

**. The cohort, respondent, and judge are shared; internal computation differs.

At K=4, an eight-turn PAUSE dialogue generates 32 candidates and performs up to 64 candidate–scorer evaluations, which can be batched. We will report this computation structure alongside the performance comparisons. Latency and monetary cost have not been measured, so the current evidence establishes a performance benefit rather than a quantified cost advantage.

Why does the 14B scorer perform worse?

The component results help explain why model size alone does not predict the downstream outcome. The 14B appropriateness scorer has lower SPoRC MRR than 8B, 0.7536 versus 0.7920, despite slightly higher YIELD utility AUROC, 0.7466 versus 0.7429. The rollouts also reuse one threshold across backbones. These observations suggest ranking transfer and score calibration as possible contributors; their individual causal effects have not been isolated. We will report component quality and downstream performance together.

What supervision is needed for open-ended use?

Inference requires context and candidate questions, without ground-truth appropriateness or utility labels. Training uses observed versus mismatched dialogue contexts and outcome-derived novelty targets, reducing the need to manually label target-domain questions. NewsInterview already tests online generation of natural-language questions.

Label-free inference and reliable domain transfer are separate questions. A preliminary NI-only-to-SPoRC run yields coverage of 0.0148, demonstrating that successful transfer cannot be assumed. The rubric-based evaluation also does not establish a question-level preference advantage. We will therefore ground the demonstrated benefit in reference-information coverage and selective clarification, and identify domain transfer and unrestricted open-ended quality as further validation targets.

Metric definitions and presentation

MRR is mean reciprocal rank: MRR = 1N∑iN1ri\frac{1}{N} \sum_i^N \frac{1}{r_i}, where rᵢ is the observed response's rank in candidate group i. Calibration UAR divides unwanted asks by no-clarification-needed topics; the table's UA divides them by all evaluated topics, so one unwanted ask among 50 topics gives UA = 0.02. We will define these denominators at first use, add coverage uncertainty bands from a consistent turn-wise evaluation source, and compact the figure and extended tables.

Thank you again for your time and consideration. We would greatly appreciate it if you could let us know whether our responses and additional experiments have sufficiently addressed your concerns. Please feel free to let us know if any further clarification is needed. We would be happy to discuss any remaining questions.

I thank the authors for the reply. However, I'm still not convinced that the benefit justifies the added complexity, especially when the desired behavior may be already present on current models on natural dialogue, as reviewer grz3 points out as well. It also seems like much of PAUSE's gain is recovered by a simple repetition filter, at least in the present simulated setting.

Is the 0.5055 result of the frozen base generator with the learned-utility reranking also gated by the appropriateness score? The simplest baseline for this would be to have a frozen generator (the same model used for dialogue), with interventions gated by the appropriateness score and ranked by the utility function. I'm still not convinced of the necessity of training a generator.

Finally, I also found the rebuttal hard to parse. I'm keeping my score for now.

Reviewer_C4dTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is about information-seeking dialogue agents, such as LLM interviewers or search systems that ask clarifying questions. It argues that choosing the next question should depend not only on how much new information the question may bring, but also on whether it fits the current point in the conversation. A useful question can still come too early. The proposed method, PAUSE, trains two separate scorers, one for appropriateness, learned from podcast transcripts, and one for utility, learned from information-elicitation dialogues. It first filters out candidate questions that don’t fit and then picks the most useful one among the rest. The same rule is used to build SFT and DPO training data for a Qwen 8B question generator. In simulated NewsInterview interviews, coverage of the reference information is increased. On ClariQ the model asks fewer unnecessary clarification questions, but the retrieval gain over utility only selection is not significant.

Critical strengths
  • The main idea is simple and interesting. The paper asks whether a question fits the current point in the conversation is a separate decision from how informative it is. Treating fit as a filter before ranking by utility is easy to add to existing question-selection systems. The SPoRC check in Appendix A.1 is a good test of what the appropriateness scorer actually learns. The real host question scores high at its real position and low when moved earlier, while a matched decoy scores low at both positions. This shows the scorer reacts to timing, not just topic, which is nice.

  • The main NewsInterview comparison is fairly well controlled. The gain over utility-only selection holds across three scorer backbones (and significant for two of them) and when the judge model is swapped, and PAUSE beats the prompt-only LLM interviewers. So the improvement looks real, even if what causes it is still unclear.

Critical weaknesses
  • The paper doesn’t show that the NewsInterview gain comes from better timing. The main claim is that some useful questions are premature and should be pushed back. But the paper doesn’t report which candidates the appropriateness filter removes, how often it changes the selected question or whether the removed questions were actually premature rather than off-topic, ungrounded or badly written.

  • Evaluation is fully simulated, with no human check of whether the chosen questions are better timed. Without this evidence, the results show that a second learned score helps, not that timing matters, which is the central claim of the paper.

  • The ClariQ results add little, and the reported configuration was chosen on the test set. The test set is only 50 topics. PAUSE’s gain over utility-only selection is +0.008 [−0.005, 0.026], which is not significant, and the drop in unwanted questions is from 4 topics to 1 (if I got it correctly).

Questions for authors

See weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer C4dT

Thank you for recognizing the value of separating contextual fit from information utility and the controlled NewsInterview evaluation. We added targeted controls to examine what drives the coverage gain and to clarify the evidence for selective clarification.

What drives the NewsInterview gain?

Adding the gate improves coverage even when the utility scorer receives the already-discussed state: the paired Coverage@8 gain is **+0.0349

00˙015,00˙6830\.0015, 0\.0683

**. This shows that the gate contributes beyond explicitly providing dialogue history to the tested utility scorer.

The deduplication control tests a simpler selection alternative. A fixed n-gram filter removes 17.8% of candidates, falls back on 8.5% of turns, and improves coverage by +0.0517

00˙166,00˙8690\.0166, 0\.0869

, reaching 0.5560. PAUSE reaches 0.5606. This establishes lexical deduplication as an effective baseline, while the history-aware control provides separate evidence for the value of adding contextual filtering.

The SPoRC probe tests placement sensitivity by scoring recorded host responses at observed and earlier positions alongside matched decoys. The reported example-clustered response–position interaction is **+0.558

04˙65,06˙460\.465, 0\.646

**. We will distinguish this scorer-level evidence from the downstream mechanism: the coverage experiments establish a benefit from contextual selection, but do not quantify how much of that benefit specifically comes from correcting premature questions.

How reliable is the coverage improvement?

The gain replicates across all three tested generation seeds and all three coverage judges. With the generator and eight-question budget fixed, Coverage@8 rises from 0.5042 to 0.5606 on 348 conversations, a paired gain of +0.0564

00˙216,00˙9120\.0216, 0\.0912

(95% bootstrap CI). The three seed-specific gains are +0.0564, +0.0398, and +0.0559, each with an interval excluding zero; the conversation-clustered pooled gain is +0.0507

00˙310,00˙7160\.0310, 0\.0716

. Qwen3.5, DeepSeek-V4, and GLM-5.3-Flash judges yield gains of +0.0564, +0.0547, and +0.0441, respectively, again with all intervals excluding zero.

These checks strengthen the automated information-recovery result. Human timing evaluation and a completed candidate-level fit analysis remain outstanding, so we will state the demonstrated outcome as improved coverage under contextual selection and keep the timing-specific mechanism question separate.

What does ClariQ establish, and how was the configuration selected?

The additional ClariQ result supports selective question use with a retrieval gain over NoAsk. On 50 test topics, the development-frozen T+V2 configuration improves nDCG@20 over NoAsk by **+0.0364

00˙056,00˙7130\.0056, 0\.0713

**. Compared with the matched V2 always-ask policy, it asks on 42 rather than 50 topics, and unwanted asks decrease from 4 to 1. The matched retrieval difference is −0.0026

−00˙200,00˙179−0\.0200, 0\.0179

, indicating no statistically detectable change in retrieval in this sample, rather than establishing equivalence.

We will center the ClariQ discussion on this retrieval–question-use trade-off and report counts alongside rates. To address configuration selection, we will label the original grid as exploratory and separately document the full configuration and threshold-selection timeline for the additional evaluation. A development-selected threshold alone does not establish that every model choice preceded test evaluation.

Thank you again for your time and consideration. We would greatly appreciate it if you could let us know whether our responses and additional experiments have sufficiently addressed your concerns. Please feel free to let us know if any further clarification is needed. We would be happy to discuss any remaining questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.