acceptodds
Under review as a conference paper at ICLR 2027

Asking for What Was Never Requested: Horizontal and Vertical Proactivity in Agents

17 views

Abstract

An agent that uses tools typically responds to what the user explicitly asks, yet completing the task may require information the user never requested. Work on proactive agents mainly studies whether and when an agent should act on its own, not what information it should pursue. We study a distinct axis of proactivity: its content. Horizontal proactivity pursues unstated information that the current context already identifies, and vertical proactivity pursues needs that only earlier evidence reveals. A need graph, recovered from a benchmark's own decomposition, records which needs depend on which, so both forms, and whether the agent stops at the right time, can be scored from a transcript without a model judge. To learn this behavior, we propose Q&D (questioner and drafter), which trains a questioner to prefer the question whose continuation retrieves more of the required evidence, with no reward model or judge. On held-out splits of three multi-hop question-answering benchmarks, at equal retrieval spend, the trained questioner improves both forms of proactivity over the same model, prompted, and outperforms a prompted model 15× larger in the same role on two of the three, and the gain persists after controlling for question volume and length. Without further training, we place the questioner in an interactive customer-service agent with a simulated customer, where it completes more tasks while asking fewer questions, and in retail it outperforms the 15× larger model with fewer follow-up turns from the customer. These results show that proactivity depends not only on whether an agent acts without being asked, but also on what it chooses to pursue and when it stops.

Primary area:
applications to robotics, autonomy, planning
Keywords:
proactive agents, question asking, information seeking, LLM agents, retrieval-augmented generation, multi-hop reasoning, preference optimization, tool use

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ddWCOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how AI agents can be proactively helpful by seeking more necessary information on their own, rather than just reacting to explicit user prompts. The authors introduce two types of proactivity: horizontal (finding obvious missing information from the start) and vertical (following a trail where newly discovered evidence reveals the need to search for another). They use a need graph that maps out all the required evidence for a specific task. Then, the authors also propose Q&D, a new training method that teaches a smaller model to prioritize asking questions that uncover the most useful evidence. The results show that this approach helps the AI find deeper information, even allowing it to outperform a model 15x its size. Furthermore, when applied to a simulated customer service setting, the agent successfully solved more user problems.

Critical strengths
  • The paper introduces horizontal and vertical proactivity and operationalizes them through need graphs and measurable metrics. This provides a structured way to evaluate several things: whether agents proactively seek information, what information it finds, and how it follows dependencies. The framework is a meaningful contribution because it enables more fine-grained, judge-free evaluation of behavior seeking information.
  • The Q&D method consistently improves required-evidence coverage over the same prompted model across three multi-hop QA benchmarks. The experiments also include comparisons with larger models and ablations that help isolate the effect of training. Moreover, the improvements in retail task success on τ2{\tau}^2-bench, without additional training, provide encouraging evidence that the learned behavior can transfer beyond the original QA setting.
Critical weaknesses
  • The evaluation relies on small, fixed retrieval pools where most prerequisite edges do not meaningfully constrain retrieval. Although the authors acknowledge and investigate this issue, the results mainly demonstrate better coverage of deeper decomposition steps rather than dependency-driven retrieval in more realistic settings.
  • While the coverage of evidence is gained consistently, the improvements in QA (Table 2) remain modest and statistically inconclusive. Stopping behavior is also imperfect, particularly on StrategyQA, where the questioner often stops before collecting all required evidence. This raises questions about how effectively retrieval gains translate into end-to-end task performance.
  • The comparisons with recent learning-based adaptive retrieval and stopping methods are quite limited. Furthermore, the authors report only two final-stage training seeds and acknowledge deviations from the original checkpoint selection rule. These factors limit confidence in the robustness and generality of the findings.
Questions for authors

Please address the weaknesses above. There are several questions starting from the weaknesses. Authors can treat them as suggested questions to answer during the rebuttal phase if they find them useful:

  • Appendix I.4 provides a useful analysis of prerequisite dependencies. Could the authors elaborate on how they expect Q&D's vertical-proactivity gains to generalize to larger retrieval corpora, where following dependencies may be more important? Are there any observations from the current experiments that support this expectation?
  • What do the authors see as the main bottleneck of translate evidence coverage into better answer quality? Can it be pẻmature stopping, the use of retrieved evidence by the answerer, or other factors?
  • Could the authors clarify how Q&D differs from recent learning-based adaptive retrieval and stopping approaches, particularly in terms of training signals and decision-making? Additionally, given the shared training components and checkpoint selection limitations noted in Section 7, how should readers interpret the robustness of the reported improvements across training runs?
6 comments
AuthorsOpenReview ↗

Q1 comment

Thank you for finding the paper "clear and easy to follow" and for recognizing our need-graph framework as a "meaningful contribution" enabling "fine-grained, judge-free evaluation." We also appreciate your recognition of Q&D's consistent retrieval gains and its transfer beyond QA without additional training.

Your three points motivated additional informative experiments:


Q1. Do Q&D's gains hold in larger retrieval corpora where dependencies matter?

Q&D continues to recover more required evidence from larger document collections, including evidence that depends on earlier answers. Its subsequent questions also benefit from information retrieved in previous steps.

To test whether the original small document pools made retrieval too easy, we combined the pools of all 200 held-out tasks into one shared collection per benchmark, increasing their size to 1,524–3,207 paragraphs (126–160× their original size). Keeping the retriever, top-k settings, and retrieval budgets unchanged, Q&D still improves equal-spend coverage over prompting by +12.2 [8.7, 16.0], +5.4 [1.1, 10.0], and +3.1 [0.7, 5.6] points on MuSiQue, StrategyQA, and 2WikiMultiHopQA, respectively.

The gains extend to evidence whose retrieval depends on earlier answers: In the larger collections, we identified cases where providing an earlier answer was necessary to retrieve downstream evidence under our test. On these cases in MuSiQue, Q&D recovers that evidence 25.9 [6.9, 44.0] percentage points more often than prompting.

We also tested whether Q&D retrieves prerequisites before their downstream evidence, but found no clear improvement in this order. This does not necessarily contradict our objective of recovering required evidence efficiently, since one retrieval can return multiple pieces of evidence. More importantly, our existing evidence-hiding ablations (Section 5) show that preventing the questioner from seeing previously retrieved evidence reduces MuSiQue coverage by 8.4 and 8.7 points across the two seeds. This suggests that Q&D uses earlier discoveries to guide subsequent questions, even though the results do not establish that it systematically follows prerequisite chains.

We have added the larger-pool experiments and dependency analyses in Appendices E.1 and I.5.

AuthorsOpenReview ↗

Q2 Comment

Q2. Why doesn't better evidence coverage always improve answer quality?

We identify two bottlenecks between finding evidence and producing correct answers: limitations in final-answer generation and premature stopping: Q&D trains the questioner to seek missing information proactively, while a separate, fixed LLM uses the collected evidence to generate the final answer. On MuSiQue, Q&D retrieves evidence containing the answer more often than prompting (94.4% vs. 84.8%), but final-answer F1 remains similar when that evidence is available (47.8% vs. 48.0%). To examine it, we replaced the fixed answer-generating model with Claude Opus 5 (on the exact previous answerer input) and it increased Q&D's absolute F1 by 31.3%, suggesting substantial room for improving answer generation independently of information seeking. On StrategyQA, Q&D retrieves more required evidence than prompting at equal retrieval spend, but premature stopping limits this advantage to be translated to a better F1. Among runs ending without the evidence needed for the final answer, Q&D chooses to stop rather than reaching the retrieval limit in 88.5% of cases, compared with 39.0% for prompting.

Our new experiments show that improved evidence collection can lead to better final answers: In the larger document collections described in Q1, Q&D improves MuSiQue F1 over prompting by +5.8 [3.0, 9.1] points. To further examine this connection, we compared Q&D before and after its final training stage, which teaches the questioner to keep searching when evidence is missing. This training improves evidence coverage before stopping on all three benchmarks and increases answer F1 on MuSiQue (+3.3 [0.9, 6.4]) and 2WikiMultiHopQA (+3.0 [0.5, 5.9]), although it requires additional retrieval calls.

We have added these analyses in Appendices G.1 and D.1 and clarified the findings in Section 5.

AuthorsOpenReview ↗

Q3 Comment

Q3. How does Q&D differ from learned retrieval methods, and how robust are its gains?

Q&D learns what information to seek next and when to stop, using the evidence recovered by its questions as the training signal: This differs from Adaptive-RAG, which learns to select a retrieval strategy based on question complexity, and Search-R1, which learns search-and-answer decisions from final-answer rewards. Q&D instead compares candidate questions by the required evidence their subsequent retrievals recover, and trains the questioner to continue when evidence is still missing. Need graphs identify the required evidence during training but are not provided at inference time.

Compared with Search-R1, Q&D recovers more evidence and produces better answers when each method decides when to stop, although results at equal retrieval cost are mixed: Our paper already compares Q&D with PAR²-RAG, a structured retrieval approach that does not require training (Section 5). To address your concern about learned methods, we additionally evaluated the released Search-R1 7B on the same held-out tasks, keeping the retriever, document collections, answer-generation components, and retrieval limits fixed (Appendix E.2). When each method chooses when to stop, Q&D recovers more required evidence on all three benchmarks, improving answer F1 by +13.1, +4.2, and +11.5 percentage points on MuSiQue, StrategyQA, and 2WikiMultiHopQA, respectively (all statistically significant). Search-R1 generally stops earlier and uses fewer retrieval calls, which helps explain the difference. At equal retrieval cost, Q&D leads on MuSiQue, performs similarly on StrategyQA, and trails on 2WikiMultiHopQA. Since the released models also differ in their base models and training data, this comparison does not isolate the effect of their training objectives.

The coverage improvements are consistent across additional training runs, but these do not establish robustness of the full training pipeline: We extended the original two final-stage training runs to four. All four improve coverage over prompting on all three benchmarks, with all 12 confidence intervals excluding zero and only 0.4–0.7 percentage points of variation across runs. However, these runs share the same earlier checkpoint and training data, so they establish reproducibility of the final training stage rather than independent replication of the complete pipeline. No checkpoint satisfied all the original selection criteria, including the stopping requirement. We now explain the alternative selection procedure and its limitations in Section 7 and Appendix L.

Reviewer_dyU6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper distinguishes horizontal expansion of information needs from vertical discovery of needs revealed by earlier evidence. Q&D trains a Qwen3-8B questioner using question outcomes and stopping behavior, with a frozen GPT-OSS-120B drafter/answerer. Evaluation uses requirement graphs, multi-hop QA, diagnostic audits, and service-task continuations from intermediate states.

Critical strengths
  • Retrieval gains have useful controls. At equal retrieval spending, evidence coverage improves by 11.2, 7.0, and 5.0 percentage points on three 200-task test sets, with confidence intervals above zero. Budget curves, length controls, and evidence-blind analyses suggest that longer retrieval alone does not explain the gains.
  • Transfer evidence extends beyond graph metrics. Retail continuations improve success by about 20 percentage points over the same 8B prompting baseline. This supports practical value, within the limited setting of 34 continuation states per domain and a retained 120B drafter/answerer.
Critical weaknesses
  • The metrics do not clearly distinguish proactivity from evidence recall. Of 1,040 probed dependency edges, 20 are validated, 654 refuted, and 366 untestable. The top-5 probe differs from some main settings, but MuSiQue also uses top-5 and validates only 12/466 edges. Reaching deeper graph nodes therefore does not establish that earlier evidence was needed to discover those requirements.
  • The final stopping stage's added value is unclear. The question-pair row in Table 3 is not its actual starting checkpoint, and Table 9 lacks a direct paired comparison with that checkpoint. Stop-when-done is about 43%/55%, below the predefined 85% target; answer-F1 difference intervals include zero on all three main test sets. These results leave the overall acquisition–stopping–answer trade-off uncertain, rather than demonstrating no answer benefit.
Questions for authors
  • Metric validity: What does the dependency audit show at each main experiment's actual top-k? Please distinguish ordered acquisition along validated dependencies from ordinary coverage, and limit the interpretation where evidence for vertical discovery is insufficient.
  • Training-stage analysis: Please directly compare the final stage's actual starting checkpoint with the final runs on the same tasks and budgets, reporting paired differences and uncertainty for coverage, stopping, and answer quality.
7 comments
AuthorsOpenReview ↗

Q1 Comment

Thank you for recognizing the "useful controls" supporting our equal-spend retrieval gains and the improvement in retail task success that "extends beyond graph metrics".

We ran additional experiments to address your two points.


Q1. Does graph depth demonstrate dependency-driven discovery?

Graph depth measures downstream evidence coverage, but does not establish how the agent discovered it: Following your suggestion, we examined whether the benchmark dependencies actually constrain retrieval, whether Q&D's gains hold under harder retrieval conditions, and whether it uses earlier discoveries.

Most testable dependencies do not require retrieving prerequisites first: We repeated the dependency audit at the actual retrieval settings (k = 5, 2, and 3). Of 1,040 edges, 27 satisfy the strict retrieval-dependency test (vs. 20 originally), 589 are refuted, and 424 remain inconclusive. MuSiQue remains at 12 validated edges out of 466. This means that reaching deeper evidence does not necessarily require retrieving its prerequisites first, so depth alone cannot establish dependency-driven discovery.

Q&D's advantage holds under harder retrieval conditions: We expanded the document pools to 1,524–3,207 paragraphs per benchmark (126–160× the original pool size). At equal retrieval spend, Q&D still improves coverage over prompting by +12.2 [8.7, 16.0], +5.4 [1.1, 10.0], and +3.1 [0.7, 5.6] points on MuSiQue, StrategyQA, and 2WikiMultiHopQA, respectively. The larger pools also yielded 71 validated retrieval dependencies. On the 31 downstream needs in MuSiQue with validated dependencies, Q&D recovers the required evidence 25.9 [6.9, 44.0] points more often than prompting.

Q&D benefits from earlier evidence, even without following a fixed retrieval order: We tested whether Q&D retrieves prerequisites before downstream evidence, including on validated dependencies, but found no clear improvement over prompting. Following a prescribed order is not our objective, since a single retrieval can recover multiple pieces of evidence. Instead, we aim to use earlier discoveries to guide subsequent questions and recover the required evidence efficiently. Our existing evidence-hiding ablations (Section 5) support this: preventing the questioner from seeing previously retrieved evidence reduces MuSiQue coverage by 8.4 and 8.7 points across the two training seeds. This shows that access to earlier evidence helps Q&D decide what to ask next and recover more of the required information.

We have clarified the interpretation of our depth metrics and added the new dependency and ordering analyses in Appendices I.5 and E.1.

AuthorsOpenReview ↗

Q2 Comment

Q2. What does the final training stage contribute?

The final stage reduces premature stopping, improving evidence coverage on all three benchmarks and answer F1 on two, with a trade-off in stopping efficiency:

As you pointed out, the question-pair row in Table 3 is a separate ablation trained from a pooled imitation checkpoint, not the final stage's actual starting checkpoint. The actual checkpoint was already trained to prefer more useful questions, and the final stage adds training to favor asking over stopping when evidence is missing. We therefore directly compared the model before and after this stage on the same 200 held-out tasks per benchmark, with up to eight retrieval calls.

The table reports final stage minus its starting checkpoint, with paired 95% confidence intervals. Differences are in percentage points, except for questions per task row.

MetricMuSiQueStrategyQA2Wiki
Coverage, at own stop+8.9 [6.8, 11.5]+9.3 [7.1, 11.9]+4.6 [2.5, 7.5]
Answer F1, at own stop+3.3 [0.9, 6.4]+2.1 [−0.6, 5.3]+3.0 [0.5, 5.9]
Ask while evidence is missing+9.9 [8.2, 11.8]+13.5 [10.9, 16.0]+4.8 [3.1, 6.8]
Stop once evidence is complete−48.3 [−54.1, −41.9]−51.8 [−58.7, −45.2]−17.0 [−26.7, −10.7]
Questions per task+1.71 [1.51, 1.95]+1.55 [1.31, 1.83]+0.39 [0.25, 0.58]
Coverage, equal spend−1.8 [−4.0, 0.6]+1.8 [−0.4, 4.5]−0.4 [−3.0, 1.6]

The final stage's main contribution is learning when to keep asking: It asks more often when evidence is missing and consequently recovers more evidence before stopping. At equal retrieval spend, we find no detectable coverage difference between these two trained checkpoints, which indicates that the starting checkpoint had already learned to prioritize useful questions.

The additional evidence also improves answer quality: Relative to its starting checkpoint, the final stage improves answer F1 on MuSiQue (+3.3 [0.9, 6.4]) and 2WikiMultiHopQA (+3.0 [0.5, 5.9]). The original F1 comparisons against prompting remain inconclusive on the original retrieval pools. However, in the larger-pool MuSiQue experiment described in Q1, Q&D also improves F1 over prompting by +5.8 [3.0, 9.1] points.

The trade-off is stopping efficiency: The final stage also continues asking after all required evidence has been found, leaving stop-when-done below our 85% development criterion. Thus, it improves evidence acquisition and answer quality (on 2/3 benchmarks), while balancing persistence with timely stopping remains a challenge, as we acknowledged.

We have added the direct checkpoint comparison in Appendix D.1 and clarified this trade-off in the revised paper.


We hope these additional experiments address your concerns and would appreciate a reconsideration of your rating.

AuthorsOpenReview ↗

Discussion period approaches its end

Dear Reviewer,

As the discussion period approaches its end, we wanted to briefly follow up on our detailed responses to your concerns.

We conducted additional experiments addressing both of your main questions and incorporated the results into the uploaded revised manuscript.

We would appreciate your feedback. If it addressed your points, please consider updating your score. If any questions still remains, we would be happy to discuss them further.

Thank you for your time and consideration.

Reviewer_pH2LOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript focuses on developing proactive search agents, with an exploratory proof-of-concept experiment in the downstream application in customer service agents. The authors formulate proactivity along two axes: vertical proactivity, going deeper into other needs that become available after a response to the current need/question is obtained, and horizontal proactivity, collecting adjacent information that is not yet important for the current state of the search.

Critical strengths
  • Proactive search, being a difficult task, is well formulated within the structure of the needs graph. Using coverage of intermediate concepts (needs) necessary to obtain the final answer/resolve a task is an essential first step to a grounded evaluation of search agents and is the main strength of this work.
  • The experiments conducted in the work are quite extensive and well designed to test the proposed hypothesis. The only issue is the limited exploration of open-weight models, with only one model (GPT-OSS-120B) used as the untrained baseline, which makes it hard to generalize the findings. I would encourage the authors to also explore other similarly sized models, such as gemma-4-12B (though not necessary, it would make the findings richer).
Critical weaknesses
  • The generalizability of the proposed approach (i.e., training a questioner to maximize coverage) that is evaluated to maximize the coverage of the intermediate needs is limited to 25 retail and 17 airline tasks from the tau^2 benchmark. While it does provide some evidence, it is still limited in its ability to generalize the efficacy of the proposed Q&D approach.
  • While this is not an ethical concern for the rigorous experiments conducted in this work, there is however, a concern about how these proposed Q&D agentic systems are inherently designed. Horizontal proactivity is defined as pursuing information the user did not mention (e.g., information about other orders). If it is an unbounded investigation, there could be ethical and privacy risks in developing such systems that might proactively collect other--best case, irrelevant; worst case, harmful--information for the downstream QA/decision-making process. If the authors agree, they should at least mention such possibilities in the ethical considerations of the manuscript.
Questions for authors
  • Did the authors consider conducting small-scale user studies to test out the model's qualities in a real-world scenario?
  • How would you extend the proactive agents to other domains, besides the multi-hop QA training pipeline, that can have generalizability concerns due to the nature of the need's graph topology variations across tasks?
2 comments
AuthorsOpenReview ↗

Open source models | Q1 | W1

Thank you for recognizing our need-graph formulation as the paper's "main strength" and describing the experiments as "quite extensive and well designed."


Additional open-weight models. Following your suggestion to "explore other similarly sized models," we evaluated three new baselines and found that Q&D retains its evidence-coverage advantage: We tested Qwen3-32B, Granite 3.3 8B, and Gemma-4-12B, including the model you suggested, as prompted questioners on the same held-out QA tasks. We kept the retriever, answer-generation components, and retrieval limits fixed. At equal retrieval cost, all nine comparisons favor Q&D, with statistically significant gains in six.

Prompted comparatorMuSiQueStrategyQA2WikiMultiHopQA
Qwen3-32B+7.4+3.6+3.2
Granite 3.3 8B+8.9+6.8+8.5
Gemma-4-12B+10.7+2.5+2.6

Coverage improvement of Q&D over each prompted comparator model, in percentage points. Bold values indicate statistically significant differences; the other three remain inconclusive. Full 95% confidence intervals are reported in Appendix J.1.

W1. Q&D improves task success beyond QA without additional training, supported by extensive evaluations, and its questioning behavior also transfers to open-ended human tasks. We agree that the number of distinct customer-service tasks is limited. However, our evaluation covers 42 of the 46 held-out retail and airline tasks, with the remaining four lacking suitable dialogue traces for continuation. Just to emphasize how long horizon each task is, across the trained and prompted questioners, both training seeds, and the larger-model comparison, we conducted 1,632 runs involving over 140,000 recorded model calls. Q&D improves retail task success by approximately 20 percentage points over prompting, while our additional airline evaluation shows gains of 6–7 points under both prompt variants. Although these repeated runs do not expand the number of distinct tasks, they provide a substantial evaluation of the observed transfer gains. To examine transfer beyond these benchmark tasks, we also evaluated Q&D on open-ended human coding sessions, as described in Q1.

Q1. Although a live user study was not feasible during the rebuttal period, our new evaluation shows that Q&D anticipates information people later request in real coding sessions: As an intermediate test, we used 34 recorded Claude Code sessions involving two consenting participants. From these logs, we extracted 1,126 intermediate conversation states at which the coding assistant had to decide its next action, and replayed the trained Q&D questioner unchanged on each state. Unlike training, this setting has no need graphs or fixed evidence pools. We then examined whether its questions anticipated requests the participant actually made within the next five turns.

Questions evaluatedMatched a later user request
Q&D questions13.7% [11.8, 15.8]
Unrelated questions from other sessions1.7%

Among Q&D's 142 anticipated requests, a three-model panel independently judged that 63 concerned information already identifiable from the available context. The panel also judged 96% of Q&D's questions relevant to the task. These results provide evidence of proactive information seeking beyond benchmark QA, although the replay did not involve users interacting with Q&D and therefore cannot establish improvements in user experience.

AuthorsOpenReview ↗

Comment Q2 | W2

Q2. Different need-graph structures require different task decompositions, but no change to Q&D's training method: A need graph describes the information required to complete a task and which needs depend on earlier answers. For a new domain, we would obtain or construct the corresponding decomposition and apply the same evidence-recovery training objective. Our QA benchmarks already contain different dependency structures, demonstrating that the method does not assume one fixed graph topology. Constructing a decomposition for a new domain requires domain-specific preparation, but not a new algorithm. Moreover, the graph is used for training and evaluation, not at inference time, as illustrated by the customer-service and coding-session transfers.

W2. We appreciate your concern about privacy and over-collection, and our new audits found no cross-customer record retrieval or sensitive requests identified by the model jury in the tested settings: We audited every retrieved record across 1,224 held-out customer-service runs and found none belonging to another customer. Separately, Claude Opus 5.5, GPT-Sol, and Gemini Pro evaluated the trained questioner's 1,095 questions from the coding sessions. Their consensus identified no requests for credentials, personal identifiers, payment or health information, or another person's data (Appendix O.2). These audits address concrete risks you raised, while not ruling out irrelevant information collection or guaranteeing safe deployment. Following your feedback, we also expanded the Ethics discussion to address access restrictions, data minimization, and auditing of retrieved information.


We hope these additional experiments and revisions address your concerns and welcome any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.