acceptodds
Under review as a conference paper at ICLR 2027

CLEAR: Learning to Clarify for Earlier Gains in Domain-Specific Interactive Retrieval

Abstract

Domain-specific document retrieval often begins with queries that leave relevance-critical information unspecified. Interactive retrieval elicits this information through clarification questions and query updates. However, clarification must adapt to the evolving dialogue, and questions must be prioritized to improve retrieval with fewer rounds. We propose Clarification Learning for Early and Accurate Retrieval (CLEAR), a framework that uses retrieval results and dialogue history to identify what to clarify, and retrieval feedback to learn question priorities. Progressive Residual Facet Discovery (PRFD) uses residual quantization to organize document contrasts, combining them with domain facts and dialogue history to identify requirement facets, or aspects of the information need that can be clarified. PRFD produces candidate facets; the policy selects a facet and generates a question. Turn-aware Hierarchical Policy Optimization (THPO) constructs a return matrix from per-round retrieval gains and uses hierarchical comparisons across facet categories and candidate facets for reinforcement learning. Across legal, medical, and patent benchmarks, the domain-specific configurations achieve higher NDCG@10 at R4 and mean NDCG@10 over R1-R4 than the compared interactive baselines. The patent configuration disables round-dependent prefix selection, assessing framework transfer rather than the full progressive mechanism. Legal and medical trajectories additionally quantify the proportion of queries that first reach an NDCG@10 threshold within each round budget.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Domain-Specific Retrieval, Interactive Information Retrieval, Intent Clarification, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zu4uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies interactive document retrieval in specialized domains (legal, medical, patent), where a user's initial query omits facts that determine relevance and the system asks clarification questions over several rounds with a fixed retriever.

It makes two claims. First, clarification should become progressively finer: broad questions early, fine-grained questions once the candidate documents are more focused. Second, because users may stop early, the questioning policy should be trained to obtain retrieval gains earlier, which a terminal reward does not encourage.

The proposed framework, CLEAR, has two components. Progressive Residual Facet Discovery (PRFD) encodes documents with residual quantization (RQ) codes and turns fact differences between retrieved document pairs that share an RQ prefix into candidate facets, preferring longer shared prefixes in later rounds. Turn-aware Hierarchical Policy Optimization (THPO) trains a Qwen3-4B questioning policy with stopping-weighted reward-to-go returns and advantages that combine trajectory-, facet-category- and facet-level comparisons.

Critical strengths

1. A training signal derived from the stated objective. The stopping-weighted return is derived rather than asserted. Under a stopping distribution, the first decision's return equals expected quality at stopping minus the initial quality (Eq. 13), and earlier gains are excluded from later decisions with a clear argument (Eq. 16). This ties the learning signal directly to the "earlier gains" goal.

2. Broad and transparent evaluation. The evaluation is broad:

  • three domains and seven adapted baselines, with a fixed retriever and frozen candidate pools;
  • query-paired bootstrap intervals (Fig. 5, Fig. 7);
  • unselected success and failure trajectories (Fig. 8);
  • limitations stated explicitly (App. A.3).
Critical weaknesses

1. The paper claims coarse-to-fine clarification, but neither its mechanism nor its experiments show that questions actually move from coarse to fine. Progressive granularity is the paper's first contribution (Sec. 2). It is implemented by preferring document pairs with longer shared RQ prefixes in later rounds (Sec. 3.2). A long shared prefix means two document embeddings are close; it does not mean the fact separating them is fine-grained. Take the paper's own running example (Fig. 1): two theft cases with nearly identical narratives may differ only in whether a threat was used. That fact changes the charge and is arguably the coarsest question to ask, yet such a pair shares a long prefix and would be deferred to later rounds. The paper itself disclaims the link: "Prefix length is a structural preference, not a direct measure of question specificity" (App. A.5).

The behaviour is not measured either. Sec. 4.4 reports only NDCG@10, and no analysis checks whether later questions are finer than earlier ones. Higher retrieval quality may equally come from document-grounded candidates in general. The only direct comparison is a "descriptive cross-run comparison" (Fig. 4, 7), and the patent configuration disables the schedule yet still beats all baselines.

Finally, the premise rests on a single hedged sentence (Sec. 1: fine-grained questions asked early "may rely on unsupported assumptions"). Natural alternatives are not discussed: the inventory's own category→facet hierarchy (App. C.1), a domain taxonomy, or LLM-judged specificity. Without these, the progressive component reads as a generic compression technique applied to a new setting, which weakens the novelty claim.

2. Hierarchical advantage is unmotivated, and the gains are not shown to come from the paper's new components. Neither problem in the introduction (what to ask, gaining earlier) calls for hierarchical advantage. The paper only describes it as providing "the learning signals for policy optimization" (Sec. 1) and reads its ablation post hoc as "consistent with" its role (Sec. 4.5). It gives no reason why summing trajectory-, category- and facet-level comparisons is a better credit signal than ordinary group-relative advantage or a learned critic.

This matters for attribution. The largest component gap comes from document-grounded candidates (Table 2), an idea the paper builds on from CARQ and DocFeatures (Sec. 2). The stopping-weighted return, though well derived, is standard reward-to-go. The components that are new here, progressive prefix selection (W1) and hierarchical advantage, are the ones without a rationale. The paper therefore does not show that its own contributions, rather than inherited ideas, make CLEAR effective.

Questions for authors

Please see the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 1/2): Progressive Prefix Mechanism and Patent Results

Response to W1: Mechanism, behavioral evidence, and alternative designs for progressive clarification

Thank you for your careful reading and valuable analysis.

(1) RQ organizes document comparisons; structural similarity is not equated with semantic granularity

When two document representations are close, the fact distinguishing them is not necessarily more fine-grained. We introduce shared RQ prefixes to construct a hierarchical document-comparison neighborhood within the current retrieval results. This focuses factual comparison on document pairs that share representation structure but differ in specific facts, allowing us to discover clarification facets that help distinguish candidate documents.

Increasing the shared RQ prefix length does not directly determine the specificity of generated questions. Its main role is to select document pairs for factual comparison: raising the threshold prioritizes pairs sharing more coding levels. Our design aims to identify distinguishing factual differences among documents with similar representation structure; their contribution to retrieval is examined in the controlled comparisons below.

We appreciate the reviewer's example of whether a threat was used. Such a local factual difference can indeed be decisive for the legal characterization of a case. The schedule (1,1,2,2)(1,1,2,2) specifies the minimum shared-prefix requirement for preferred candidates at each round, rather than restricting each round to a particular code level:

(\tau_1,\tau_2,\tau_3,\tau_4)=(1,1,2,2),$$ where $p(d,d')$ is the shared-prefix length. If no candidate satisfies the current threshold, the implementation falls back to candidates that remain available. For example, a pair sharing a prefix of length 3 already satisfies threshold 1 at R1. Provided the other candidate conditions are met, its factual difference can enter the candidate set and is **not forcibly deferred because of its longer prefix**. Later rounds tighten the preferred eligibility of shorter-prefix pairs; they do not make longer-prefix pairs accessible for the first time. If the reviewer's scenario occurs at R1 or R2, our framework would regard the difference between those documents as particularly worth clarifying. The rationale is to retain a broader comparison neighborhood when little user information is available and retrieved documents may differ substantially, then favor unresolved differences between structurally closer cases as answers and retrieval results are updated. Our core objective is to investigate how questioning affects retrieval in subsequent rounds while keeping the retriever unchanged, and how selecting key facets can improve retrieval within a limited interaction budget. Progressive discovery and policy learning jointly serve this objective: discovery adapts the comparison neighborhood and supplies candidate facets, while the policy learns which facets to prioritize based on current and subsequent retrieval gains. Section 3.2 of the revised manuscript now explains the shared-prefix mechanism, minimum thresholds, and candidate fallback, including why longer-prefix pairs can enter early. Appendix B.1 gives candidate construction and ranking details, and Appendix A.7 provides a four-round interaction example. **(2) Progressive behavior, retrieval gains, and the patent configuration** Our objective is not to make questions increasingly specific, but to use shared-prefix lengths to progressively identify document pairs whose differences are most worth asking about. For patent retrieval, we have added an experiment enabling the $(1,1,2,2)$ schedule: | Patent configuration | R4 NDCG@10 | Mean NDCG@10 over R1–R4 | |---|---:|---:| | Schedule disabled | 0.2878 | 0.2757 | | Progressive schedule enabled | 0.3031 | 0.2852 | Final MAP and P@5 in this additional evaluation are 0.3410 and 0.3700, respectively. Enabling progressive scheduling improves both final and four-round mean NDCG@10 over the configuration without round-dependent scheduling. These experimental results support the schedule's role in this configuration. The patent results with progressive scheduling have been updated in Table 1. The legal comparison between progressive and fixed prefixes appears in Section 4.4, Figure 4(A). The new shared-state R3 branching experiment in Appendix A.8 (Table 11) further examines the effect of tightening the prefix threshold from the same retrieval state. The alternative-design comparisons and our response to W2 continue in Part 2/2.
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 2/2): Alternative Discovery Designs and Hierarchical Advantage

This comment continues our response to W1 from Part 1/2 and then addresses W2.

(3) Clarifying the motivation and comparing more direct alternatives

The phrase "may rely on unsupported assumptions" refers to questions that presuppose an unconfirmed event or attribute: asking about its details too early may depart from the user's actual need. For example, asking "How much compensation was paid?" before confirming that compensation occurred differs from neutrally asking "Was any compensation paid?"

We thank the reviewer for suggesting alternatives. The original manuscript included a prompt-based candidate-discovery control: an LLM proposes candidate facets from retrieved documents and dialogue, and the questioning policy selects a facet and generates a question. The revision adds three alternatives: cosine similarity of continuous representations, hierarchical K-means, and a domain facet hierarchy.

MethodCore mechanismR4 NDCG@10Mean NDCG@10 over R1–R4
CLEAR (ours)RQ document comparisons with the (1,1,2,2) schedule0.53440.4985
LLM PromptAn LLM discovers candidate facets from documents and dialogue0.44490.4452
Continuous cosine similaritySimilarity of document representations before quantization0.45330.4157
Hierarchical K-meansCluster-tree paths replace RQ paths0.43080.3984
Domain facet hierarchyA manual hierarchy orders topics to specific attributes0.42630.3974

These experimental results support RQ's role in organizing document comparisons and discovering useful clarification facets.

These alternative-method comparisons have been added to Section 4.4, Table 2, and Figure 4(B), with an accompanying analysis of RQ's role in organizing document comparisons.

Response to W2: Motivation for hierarchical advantage, credit assignment, and component contributions

(1) Connection to "what to ask" and "gaining earlier"

Hierarchical advantage is a design choice that uses the action structure of our task. Specifically, "what to ask" involves priorities at both the category and facet levels: when several candidate facts are available, the model must determine which category of information to clarify and which specific facet within that category to select. These choices can affect convergence and retrieval quality in subsequent rounds. For example, among sentencing factors, the amount involved, the means used, and compensation are distinct facets that can be queried.

Ordinary group-relative advantage can learn these choices from sampled trajectory returns. Our design additionally aggregates returns for choices sharing a category or facet. Training therefore uses both individual trajectory performance and the average performance of related choices, helping the model learn reusable questioning priorities rather than judging a facet's value solely from one sampled outcome.

The connection to earlier gains comes from the stopping-weighted return shared across all levels:

w_k=P(S\geq k).$$ Obtaining the same retrieval gain earlier benefits more possible stopping rounds. Category- and facet-level comparisons use this return with the same auxiliary constraints, and thus learn selection priorities that support retrieval under a limited interaction budget. Their roles are complementary: **the return defines which behaviors are worth learning, while hierarchical comparisons organize the training signals provided by those returns.** We have added this motivation to the "Hierarchical comparisons" paragraph in Section 3.3, explaining how category and facet returns jointly contribute to training. Appendix B.3 provides the implementation details of the comparisons and standardization at each level. **(2) Additional experiments** Following the reviewer's suggestions, we conducted ablations of the hierarchical advantage components and added a Learned critic comparison: | Variant | Advantage configuration | R4 NDCG@10 | Mean NDCG@10 over R1–R4 | |---|---|---:|---:| | Full hierarchical | Trajectory + category + facet | 0.5344 | 0.4985 | | Flat | Group-relative advantage | 0.4592 | 0.4080 | | Learned critic | Learned state-value baseline | 0.4894 | 0.4423 | | w/o facet | Trajectory + category | 0.5084 | 0.4568 | | w/o category | Trajectory + facet | 0.4997 | 0.4561 | These comparisons have been added to the "Advantage controls" section of Table 3 in Section 4.4. The accompanying text analyzes Flat advantage, Learned critic, and the variants without category or facet comparisons to support the role of the hierarchical comparison components.

Dear Reviewer zu4u,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

Reviewer_Wz6uOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CLEAR to improve retrieval quality with fewer clarification rounds. It combines PRFD for identifying clarification facets from retrieved documents with THPO for learning question priorities. Experiments on legal, medical, and patent benchmarks show improved final-round and average retrieval quality.

Critical strengths
  • S1: The paper motivates interactive clarification for queries that leave important user needs unspecified. Using facets to organize follow-up questions offers a structured way to elicit missing information and guide retrieval.
  • S2: CLEAR improves both final-round and average retrieval quality over the compared baselines across legal, medical, and patent benchmarks.
Critical weaknesses
  • W1: The method description omits important details, making the framework difficult to understand and assess. For example, concrete facet examples, the policy model’s input/output format across turns, and a worked multi-round interaction are missing, leaving unclear how facet discovery, question generation, and query updates connect.
  • W2: PRFD lacks a clear justification for residual quantization: the paper does not establish how document codes help identify useful clarification facets beyond domain facts and dialogue history, leaving its core design choice insufficiently supported.
Questions for authors
  • Q1: Could you provide a worked multi-round example showing how domain facts and the facet inventory are constructed, document contrasts yield candidate facets, and prefix preferences affect their ranking?
  • Q2: How does the supervised clarification policy perform with the same PRFD configuration before THPO training, and how much additional improvement does THPO provide?
  • Q3: Which contributes more to CLEAR’s performance: selecting the appropriate facet or generating the clarification question?
7 comments
AuthorsOpenReview ↗

Response to Reviewer 2 (Part 1/3): Method Details and Worked Multi-Round Example

Thank you for raising these questions.

Response to W1/W2/Q1

We address these questions through a complete interaction example, module roles, and comparisons of candidate-discovery methods.

(1) A complete example connecting the modules

We store structured case facts under cause, event, and outcome fields. For example, "an argument arising from a debt dispute" is a cause, "stabbing the other person with a knife" is an event, and "causing serious injury" is an outcome. For each document pair, we compare corresponding fields, such as cause, to identify differing aspects.

We also predefine legal concepts and matching rules, using keywords and regular expressions to identify relevant expressions in case descriptions. For example, "returning stolen property," "restitution," and "compensation" all map to the "restitution or compensation" facet. The fixed facet inventory gives each facet an ID, category, label, and example question.

Consider a simplified example: document A mentions "compensating for the loss and obtaining forgiveness," while document B mentions "voluntarily surrendering and obtaining forgiveness." Both involve forgiveness, but they differ in compensation and surrender.

If the user has already stated that they voluntarily surrendered but has not described whether compensation was paid, the system retains compensation as a candidate facet. The policy can then select it and ask, "Was any compensation paid for the loss afterward?" A missing mention of compensation does not imply that no compensation was paid; the user's answer establishes the actual fact. This mapping process has been added to Section 3.2; the fixed facet inventory and fact resources are described in Appendix B.1.

We use the four-round coal-sale test example, where the user initially says, "Our company has a dispute over a coal purchase."

RoundInformation known before askingSelected facet and generated questionUser answer
R1A coal-sale dispute exists.Dispute focus: Is the dispute mainly about delivery, coal quality, or payment?It is about delivery. They received our payment but did not deliver all the coal.
R2The dispute concerns delivery.Delivery status: Was nothing delivered, or was some coal delivered?They delivered one third; the rest never arrived.
R3The seller stopped supply after partial delivery.Stoppage reason: Did the seller explain why they stopped delivering?They said we had not paid the balance, so they withheld the rest.
R4The seller attributes the stoppage to the unpaid balance.Payment condition: Under the contract, was the balance due before delivery, or after full delivery and acceptance?Only after all the coal was delivered and accepted.

Before R3, the retrieved documents contain these explicitly recorded facts:

ItemCase ACase BCase C
Document factsAfter partial delivery, the seller stopped supply, citing the buyer's unpaid balance; the contract required the balance to be paid before delivery of the remaining coal.After partial delivery, the seller stopped supply, citing the buyer's unpaid balance; the contract required the balance to be paid after full delivery and acceptance.After partial delivery, the seller could not continue supply because production had ceased; payment terms were not recorded.
Extracted relevant facetsDelivery status, stoppage reason, payment conditionDelivery status, stoppage reason, payment conditionDelivery status, stoppage reason
RQ code(7,12,4,9)(7,12,4,21)(7,12,18,6)

Cases A and B have RQ codes with a shared prefix of length 3, and their difference in payment terms yields a payment-condition candidate. A–C and B–C both share prefixes of length 2, and their differing reasons for stopping supply yield a stoppage-reason candidate. Both new candidates are available to the policy. Delivery status has already been clarified at R2: "The seller delivered only one third." The system therefore does not ask about delivery status again.

In this example, the policy first asks at R3, "Did the seller explain why they stopped delivering?" After the user answers that supply stopped because the balance had not been paid, it asks at R4, "Under the contract, was the balance due before delivery, or after full delivery and acceptance?"

This process illustrates the division of roles: RQ organizes which documents to compare, document facts determine which differences can be clarified, and dialogue history and the policy determine what to ask at the current round. Both R3 and R4 use a prefix threshold of 2; changes in questions also depend on the updated history and retrieval results. This four-round example and document comparison have been added to Appendix A.7 (Tables 9–10).

The policy interface and module roles follow in Part 2/3.

AuthorsOpenReview ↗

Response to Reviewer 2 (Part 2/3): Policy Interface and the Roles of RQ, Domain Facts, and History

This comment continues our response to W1/W2/Q1 from Part 1/3.

The policy's input and output at each round

At every round, the policy receives a state in the same format: the current round, accumulated visible dialogue, previously attempted facets, and current candidate facets. The history accumulates as interaction proceeds, while candidates are reconstructed from the latest retrieval results. The following simplified R3 state omits the system prompt, candidate example questions, and other auxiliary fields. The context field is a JSON string containing the accumulated dialogue, with \n denoting line breaks between turns.

{
  "visible_state": {
    "round": 3,
    "context": "User: Our company has a dispute over a coal purchase. We want to find similar cases.\nSystem: Is the dispute mainly about delivery, coal quality, or payment?\nUser: It is about delivery. They received our payment but did not deliver all the coal.\nSystem: Was nothing delivered, or was some coal delivered?\nUser: They delivered one third; the rest never arrived.",
    "attempted_factors": [
      "sale:dispute_focus",
      "sale:delivery_status"
    ]
  },
  "candidate_actions": [
    {
      "factor_id": "sale:nonperformance_reason",
      "factor_label": "Stoppage reason"
    },
    {
      "factor_id": "sale:payment_condition",
      "factor_label": "Payment condition"
    }
  ]
}

The generative policy selects one candidate facet and outputs its ID and clarification question in the following format:

<factor>sale:nonperformance_reason</factor><question>Did the seller explain why they stopped delivering?</question>

Here, <factor> identifies the selected candidate facet, and <question> contains the question shown to the user. After the user answers that supply stopped because the balance had not been paid, the validated, informative answer is incorporated into the retrieval query. The system retrieves again and reconstructs the candidates, using the same state format in the next round. Thus, the R4 payment-condition question is determined by the updated state. This input/output interface and query-update procedure have been added to Appendix A.7.

(2) The respective roles of RQ, domain facts, and history

Domain facts provide the content for comparison, history identifies facets already attempted or resolved, and RQ organizes the document-comparison neighborhood and contributes to candidate ranking. We use residual quantization because its multilevel codes allow the system to adjust the comparison neighborhood through shared prefixes without reconstructing the coding resources for every round.

In the example, the payment-condition candidate is supported by a pair with a shared-prefix length of 3, while the stoppage-reason candidate is supported by a pair with a shared-prefix length of 2. With other scoring terms equal and a positive structural weight, the former receives a higher structural score. This score contributes to ranking during candidate discovery but does not directly determine which facet the policy ultimately selects. The policy generates the final question using the candidates and visible dialogue. The prefix threshold also prioritizes document pairs; if no candidate is available, the provider falls back to the remaining eligible candidates. All three displayed pairs satisfy both thresholds 1 and 2, so this example illustrates candidate construction. The scoring and fallback rules are described in Appendix B.1; the shared-state branching experiment in Appendix A.8 further compares R3 thresholds 1 and 2.

RQ evaluation and our responses to Q2 and Q3 follow in Part 3/3.

AuthorsOpenReview ↗

Response to Reviewer 2 (Part 3/3): RQ Evaluation, THPO Gains, and Facet Selection versus Question Generation

This comment completes our response to W1/W2/Q1 and addresses Q2 and Q3.

(3) Experimental support for the RQ design

We introduce shared RQ prefixes to construct a hierarchical document-comparison neighborhood within the current retrieval results. This focuses factual comparison on document pairs that share representation structure but differ in specific facts, allowing the system to discover clarification facets that help distinguish candidate documents.

Increasing the shared RQ prefix length does not directly determine the specificity of generated questions. Its main role is to select document pairs for factual comparison: with a longer prefix, the preferred pairs share more coding levels. Our design aims to identify distinguishing facts among documents with similar representation structure, so that the resulting questions help distinguish potentially confusable candidate documents. A smaller prefix threshold retains a broader comparison neighborhood, whereas a larger threshold focuses the comparison. The actual content and usefulness of questions also depend on factual differences, dialogue history, and policy selection. This explanation has been added to Section 3.2 of the revised manuscript.

We have added controlled comparisons of candidate-discovery methods, comparing RQ with prompt-based discovery, cosine similarity of continuous representations, hierarchical K-means, and a predefined domain facet hierarchy. The RQ configuration achieves higher R4 and four-round mean NDCG@10, providing experimental support for its role in organizing document comparisons and discovering useful clarification facets. The results and analysis are provided in Section 4.4, Table 2, and Figure 4(B) of the revised manuscript.

Q2: SFT policy performance and the additional gains from THPO

(1) The training relationship between the supervised policy and THPO

Our training consists of supervised fine-tuning and reinforcement learning. We first fine-tune Qwen3-4B with supervised learning to select a questioning target from the candidate facets and generate a clarification question. We then initialize THPO with this supervised policy and further optimize the questioning policy using multi-round retrieval feedback. Thus, THPO continues training a model that has already undergone SFT.

(2) SFT performance and additional THPO gains with the same PRFD configuration

We use the supervised policy before reinforcement learning as the SFT-only control, retain the same PRFD configuration and evaluation settings, and compare it with the final CLEAR policy. The results on the legal dataset are as follows:

PolicyR4 NDCG@10Mean NDCG@10 over R1–R4
SFT-only + PRFD0.41760.3858
SFT followed by THPO + PRFD (CLEAR)0.53440.4985
Absolute improvement+0.1168+0.1127
Relative improvement+27.97%+29.21%

With the same PRFD configuration, THPO increases R4 NDCG@10 from 0.4176 to 0.5344 and four-round mean NDCG@10 from 0.3858 to 0.4985, corresponding to relative improvements of 27.97% and 29.21%, respectively. These results show that, under this legal evaluation configuration, THPO provides substantial numerical gains beyond supervised initialization, improving both final retrieval quality and average retrieval quality throughout the interaction. The training stages are clarified in Section 4.1, and the full comparison and improvement magnitudes are reported in Appendix A.9 (Table 12).

Q3: The contributions of facet selection and question generation

The core of our task is to enable the model to select, from the current candidate facets, the questioning target that is most helpful for improving subsequent retrieval quality and reducing the number of interaction rounds required. In other words, we aim to investigate how selecting and asking about key facets can improve domain-specific retrieval while keeping the retriever unchanged. Our research therefore focuses on discovering facets worth clarifying and learning their questioning priorities based on current and subsequent retrieval gains.

Expressing the selected facet as a natural-language question primarily relies on the language model's existing generation capabilities. SFT standardizes the "facet ID + clarification question" output format and aligns the question with the selected facet. THPO then uses multi-round retrieval feedback to optimize the questioning policy, with the training signal applied to the complete facet-and-question output. Question generation serves to communicate the selected clarification target to the user. These respective roles are described in Sections 3.1–3.3.

Reviewer_aiY1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLEAR, a framework for improving domain-specific document retrieval through clarification questions. The framework combines Progressive Residual Facet Discovery (PRFD), which uses document differences and dialogue history to identify candidate clarification facets, with Turn-aware Hierarchical Policy Optimization (THPO), which learns question priorities from current and subsequent retrieval gains. The questioning policy is trained from Qwen3-4B, while the retriever remains fixed. Experiments cover legal, medical, and patent retrieval using simulated users, four clarification rounds, and fixed candidate pools of up to 100 documents. CLEAR achieves higher final and average NDCG@10 than the compared interactive baselines. Additional experiments examine discovery granularity, candidate budgets, and policy-learning components. The patent configuration does not use round-dependent prefix selection, so its results assess the broader framework rather than the complete progressive mechanism.

Critical strengths
  1. The framework combines document-grounded facet discovery with retrieval-driven policy learning to address both what to clarify and which questions to prioritize. Its emphasis on retrieval quality across interaction rounds is useful when users have a limited question budget or may stop before the final round.
  2. CLEAR improves final NDCG@10 on legal retrieval from 0.448 for the strongest compared interactive baseline to 0.534, and average NDCG@10 from 0.417 to 0.499. Improvements are also reported in the medical and patent configurations. The experiments include round-level analyses, component comparisons, and query-level uncertainty estimates, providing useful evidence on the framework’s performance.
Critical weaknesses
  1. The contribution of individual components is difficult to isolate in some experiments. Appendix C.5 states that simulator and response-validation settings differ across evaluation series, but the manuscript does not clearly identify which component comparisons keep these settings fixed. Since these settings directly affect the information added to the query, the reported differences may reflect changes in the evaluation environment as well as changes in the method.
  2. The additional value of progressive facet discovery remains unclear. The legal configuration follows a predefined prefix schedule of (1, 1, 2, 2), and Appendix A.5 notes that prefix length is not a direct measure of question specificity. Although the point estimates favor CLEAR, the confidence intervals for its differences from fixed coarse discovery include zero for both final and average NDCG@10. Stronger evidence is needed to connect the proposed progression to more useful clarification questions and additional retrieval gains.
Questions for authors
  1. Which comparisons in Tables 1–3 and Figure 4 use identical simulator and response-validation settings? Please clarify the differences across evaluation series. If key component comparisons use different settings, do the reported gains remain when these settings are held fixed?
  2. How does increasing the shared RQ prefix length affect the specificity or usefulness of the generated questions? Could the authors provide examples or quantitative evidence linking this structural change to clarification behavior, and clarify the benefit of the progressive schedule over fixed coarse discovery given the uncertainty reported in Appendix A.6?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 1 (Part 1/2): Evaluation Consistency and Prefix Mechanism

Thank you for pointing out the remaining questions about the consistency of the evaluation settings and the explanation and empirical support for progressive facet discovery.

Response to Q1/W1: Evaluation consistency

(1) Evaluation configurations and unified settings

In the previous version, Tables 1–2 used gpt-5.6-terra as the simulator, whereas the evaluations for Table 3 and Figure 4 used gpt-6-sol. The original patent configuration in Table 1 also did not enable progressive prefix scheduling.

We therefore re-evaluated Table 3 and Figure 4 using gpt-5.6-terra throughout. In these component comparisons, the answering prompt, response-validation rules, and answer-acceptance and query-update rules were all identical to those of the corresponding full CLEAR configuration. The newly added comparisons also use the same simulator. Section 4.1 of the revised manuscript describes the simulated-user settings, and Table 3 and Figure 4 in Section 4.4 have been updated with the re-evaluation results, as follows.

(2) THPO component comparisons with the unified simulator

ConfigurationR4 NDCG@10Mean NDCG@10 over R1–R4
Full THPO + progressive discovery0.53440.4985
Terminal reward0.49280.4488
Shared scalar return0.51640.4584
Immediate-gain0.50630.4574
Flat advantage0.45920.4080

Under the unified evaluation settings, full THPO achieves higher R4 and four-round mean NDCG@10 than all four variants above. These results are included in the updated Table 3 in Section 4.4 of the revised manuscript, where Flat advantage is categorized as an advantage-calculation control. The analysis following the table further discusses the roles of round-specific return allocation and subsequent gains.

(3) Prefix-configuration comparisons with the unified simulator

Prefix configurationR4 NDCG@10Mean NDCG@10 over R1–R4
Progressive (1,1,2,2)0.53440.4985
Fixed threshold 1 (1,1,1,1)0.50720.4752
Fixed threshold 2 (2,2,2,2)0.48730.4424

These prefix comparisons are included in the updated Figure 4(A) in Section 4.4 of the revised manuscript. The main text also adds an analysis of how progressive scheduling adjusts the document-comparison neighborhood as the interaction proceeds.

(4) Patent evaluation with progressive scheduling enabled

In addition, the original patent configuration in Table 1 did not enable round-dependent prefix scheduling. We re-evaluated the patent configuration with progressive scheduling enabled, again using gpt-5.6-terra, and updated the patent columns in Table 1 of the revised manuscript. The results are as follows.

MethodR4 MAPR4 P@5R4 NDCG@10Mean NDCG@10 over R1–R4
CLEAR + progressive scheduling0.34100.37000.30310.2852

With progressive scheduling enabled, CLEAR's R4 NDCG@10 and four-round mean NDCG@10 on patent retrieval remain higher than those of the interactive baselines in Table 1.

Response to Q2/W2: Effects of shared prefixes on candidate selection and retrieval gains

These questions address a central issue: does the structural preference induced by shared RQ prefixes change the clarification questions in practice, and do these changes benefit retrieval?

(1) The role of shared prefixes and their relationship to clarification questions

We introduce shared RQ prefixes to construct a hierarchical document-comparison neighborhood within the current retrieval results. This focuses factual comparison on document pairs that share representation structure but differ in specific facts, allowing the system to discover clarification facets that help distinguish candidate documents.

Increasing the shared RQ prefix length does not directly determine the specificity of generated questions. Its main role is to select document pairs for factual comparison: with a longer prefix, the preferred pairs share more coding levels. Our design aims to identify distinguishing facts among documents with similar representation structure, so that the resulting questions help distinguish potentially confusable candidate documents. A smaller prefix threshold retains a broader comparison neighborhood, whereas a larger threshold focuses the comparison. The actual content and usefulness of questions also depend on factual differences, dialogue history, and policy selection. This explanation has been added to Section 3.2 of the revised manuscript.

The worked example and shared-state experiment follow in Part 2/2.

AuthorsOpenReview ↗

Response to Reviewer 1 (Part 2/2): Worked Example and Shared-State Experiment

This comment continues our response to Reviewer 1 in Part 1/2.

(2) A multi-round interaction and document-comparison example

We have added a multi-round test example to explain how shared RQ prefixes affect clarification behavior through candidate discovery.

The user's initial query is: "Our company has a dispute over a coal purchase. We want to find similar cases."

RoundInformation known before askingClarification questionUser answer
R1A coal-sale dispute exists.Is the dispute mainly about delivery, coal quality, or payment?It is about delivery. They received our payment but did not deliver all the coal.
R2The dispute concerns delivery.Was nothing delivered, or was some coal delivered?They delivered one third; the rest never arrived.
R3Supply stopped after partial delivery.Did the seller explain why they stopped delivering?They said we had not paid the balance, so they withheld the rest.
R4The stoppage involves a balance-payment dispute.Under the contract, was the balance due before delivery, or after full delivery and acceptance?Only after all the coal was delivered and accepted.

Immediately before R3, the current retrieval results contain the following documents with explicitly recorded facts.

ItemCase ACase BCase C
Document factsAfter partial delivery, the seller stopped supply, citing the buyer's unpaid balance; the contract required the balance to be paid before delivery of the remaining coal.After partial delivery, the seller stopped supply, citing the buyer's unpaid balance; the contract required the balance to be paid after full delivery and acceptance.After partial delivery, the seller could not continue supply because production had ceased; payment terms were not recorded.
Extracted relevant facetsDelivery status, stoppage reason, payment conditionDelivery status, stoppage reason, payment conditionDelivery status, stoppage reason
RQ code(7,12,4,9)(7,12,4,21)(7,12,18,6)

Cases A and B share a prefix of length 3, and their supported difference in payment terms yields a "payment-condition" candidate. A–C and B–C share prefixes of length 2, and their factual differences yield a "stoppage-reason" candidate. Delivery status has already been clarified at R2 and is therefore excluded; stoppage reason and payment condition remain available for selection.

At R3, the policy selects the stoppage reason and asks, "Did the seller explain why they stopped delivering?" After the user answers that supply stopped because the balance had not been paid, the system updates the query and retrieves again. At R4, it asks, "Under the contract, was the balance due before delivery, or after full delivery and acceptance?" Both R3 and R4 use a threshold of 2, showing that changes in questions also depend on the updated dialogue and retrieval results.

This process illustrates the division of roles: RQ organizes which documents to compare, document facts determine which differences can be clarified, and dialogue history and the policy determine what to ask at the current round. Code IDs themselves do not correspond to specific legal facts. This four-round example, together with the document facts and RQ codes, has been added to Appendix A.7 of the revised manuscript.

(3) A shared-state branching experiment

To further examine the role of progressive discovery relative to a fixed prefix threshold of 1, we have added a shared-state branching experiment. We hold fixed the question–answer history of the first two rounds and the retrieved documents and their ranking at the end of R2, and generate R3 questions separately from this common R2 state. The fixed-threshold branch retains threshold 1, while the progressive branch raises it to 2.

Both branches use the same gpt-5.6-terra simulator and answering prompt. When both the question and context are identical, the branches share one newly generated answer. The results are as follows.

MetricFixed: R3 threshold 1Progressive: R3 threshold 2
Shared starting R2 NDCG@100.46960.4696
R3 NDCG@10 after answering0.48770.5148
R2-to-R3 NDCG@10 gain0.01810.0452
Mean candidate-facet count11.22506.7875

Starting from the same R2 state, the progressive branch achieves an R3 NDCG@10 that is 0.0271 higher than the branch with fixed threshold 1, while the mean candidate-facet count decreases from 11.2250 to 6.7875. This provides additional evidence for the role of tightening the comparison neighborhood in later rounds. The experiment and its analysis have been added to Appendix A.8 (Table 11) of the revised manuscript. In light of the confidence intervals including zero in the original comparison, we restrict this conclusion to the observed gains under the shared-state R3 intervention.

Dear Reviewer aiY1,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.