acceptodds
Under review as a conference paper at ICLR 2027

Found but Not Written: Understanding and Reducing Evidence Attrition in Deep Research

Abstract

Deep Research agents can retrieve relevant evidence yet omit details readers need. We study this evidence-to-answer attrition on 132 bilingual DeepResearch-Bench II (DRB2) tasks, where a selector flags candidate omissions in about 80% of reports. In controlled completion experiments, targeted additions fully express more omitted findings than untargeted additions, without improving mean report quality. We introduce Evidence-Grounded Residual Completion (EGRC), which organizes the answer around the request and revisits every original omission to retain content already expressed, revise a passage, or justify deferral. On 36 DRB2 test tasks, requiring a decision for every omission adds 3.9 quality points over freely selecting from the same omission list, with the same organized starting answer and nearly equal final length. The full method improves quality by 4.7 and 5.0 points over untargeted completion and two-pass self-checking, respectively, while recovering more omissions. On 23 scholarly tasks constructed following DeepScholar-Bench, EGRC improves recovery by 33 percentage points and reference-content coverage by 8.8 points over matched autonomous revision at nearly equal length. The content advantage persists under report-only assessment, and blinded readers favor EGRC over completion baselines. Explicit finding-level decisions improve recovery and report quality beyond answer organization and access to an omission list.

Primary area:
foundation or frontier models, including LLMs
Keywords:
deep research agents, evidence utilization, grounded generation, long-form report revision

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tCrNOpenReview ↗
Rating
3(+2)
Confidence
3(+1)
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors proposed a deep research framework, Evidence-Grounded Residual Completion (EGRC), which is claimed to recover information omitted from the findings and produce a better report. Experiments were conducted on DRB2, and the proposed method achieved higher scores for "quality" and "recovery."

Critical strengths
  • The paper proposed a method that collects omitted information from findings to improve the quality of the report.

  • The experimental results include human annotation that compares the proposed method with baseline(s).

Critical weaknesses
  • The abstract and introduction are terrible. A lot of undefined terms (e.g., targeted/untargeted completion, omission, selector, flag, etc.) appear throughout, making it very difficult to understand the paper's main claim and the task the authors attempted to focus on.

  • The methodology section (Section 3 & 4) missed important details. (Please see questions for the details.)

  • The experiment involved designs that are not justified. Metrics are not well defined. The generalizability and practicality of EGRC were not addressed.

Questions for authors

Presentation

  • The paper is hard to follow. Terms without definitions appear early in the paper, (e.g., targeted/untargeted completion, omission, selector). What's worse is that after reading the paper, I am still uncertain about what these terms mean.

Methodology

  • The methodology part is handwavy. The authors mentioned a selector proposes candidate omissions using the "request, findings, and report" (line 143), but how? From the latter part of the paper (Section 5), the selectors seem to be generative models, so how to prove the models will not omit useful information, as if they did when proposing the initial draft, and lead to false positives?

  • How to justify the correctness of M(t,e) (line 147)?

Experiment

  • The most frequently used metrics in the paper, Quality and Recovery are not defined properly. I suppose quality is DRB2’s overall score, but the authors need to include a definition for the uncommon metrics before discussing experimental results based on the metrics.

  • The experiment did not contain results using different generative models as backbones. Thus, it is unclear whether EGRC is generalizable.

  • Another important factor for deep research is the cost. Considering EGRC needs a selector to validate all of the "findings", the token usage is probably larger than the baseline in the paper. Considering the improvement is not huge on "Quality", it is necessary to provide the cost (as token usage or converted to USD) to better understand the practicality of EGRC.

13 comments
AuthorsOpenReview ↗

Responses to Reviewer tCrN (1/6)

Dear Reviewer tCrN,

Thank you for reviewing our paper and for recognizing the human evaluation as a strength. Below, we clarify the task, method, and evaluation, including the definitions and experimental evidence already provided in the original submission. We also present an additional experiment with GPT as the writer and a more complete cost analysis. We have revised the manuscript to make these explanations easier to follow.

Responses to weaknesses and questions

1. Task and terminology (W1; Presentation question).

We study how Deep Research agents use acquired evidence when writing long-form research reports. An agent may have found the information needed to answer a user's request, yet leave important details out of the report. We call this loss evidence-to-answer attrition. EGRC revises the report to recover useful missing details and improve the answer to the user's request. It uses the evidence already collected, without conducting another search.

In the illustrative policy example in Figure 1, a source document retrieved by the agent states that a policy applies to firms with fewer than 50 employees and excludes subsidiaries. A report saying only that eligibility depends on firm size is supported, but it does not tell the reader which firms qualify. A finding is a specific statement from a source, including the conditions needed to understand it. We prompt a large language model (LLM) to compare the acquired findings with the initial report and identify relevant information that is not fully expressed. We call this model the selector. Its selections are candidate omissions, and “flag” simply means marking a finding as a candidate omission.

The paper combines an analysis of evidence-to-answer attrition with a method for reducing it. Section 3 examines how often reports omit acquired findings and uses completion experiments to study what the writer selects and fully expresses. For these experiments, the original report stays unchanged. In untargeted completion (Generic), the writing model receives summaries of all acquired findings and chooses useful additions. In targeted completion, it receives the same summaries plus the omission list and is asked to address the listed findings. Both use the same acquired findings, writing model, and addition budget. The Two-pass baseline first asks the writer to find its own omissions and then writes additions in a separate call.

Based on this analysis, EGRC (Section 4) organizes the answer around the user's request and requires the writer to make a decision about every original omission. Source-reviewed local edits integrate the missing details into the relevant passages. Section 5 evaluates EGRC's effects on recovery and report quality. It also examines the contribution of its key design choices and tests the method across writing models and research tasks.

We have revised the abstract to describe what may be missing or incompletely expressed in a report, how additions are targeted, and what the writer does during revision. The introduction explains the main terms before discussing the completion results, and Section 3.3 introduces the writing conditions before comparing them. These changes clarify the connection between the problem analysis, the revision method, and its evaluation.

AuthorsOpenReview ↗

Responses to Reviewer tCrN (2/6)

2. Omission selection (W2; Methodology question 1).

Section 3.1 of the original submission defines candidate omissions and specifies the selector's inputs: the request, acquired findings, and initial report.

We prompt an LLM to act as the selector. It receives the user's request, the complete initial report, and summaries of all acquired findings, with their identifiers and source links. The prompt asks which findings contain useful information that the report has not fully expressed, including missing facts, corrections, quantities, comparisons, and qualifications needed to answer the request. The model returns finding identifiers, not another report. The program checks that every identifier is in the supplied collection and appears only once. These selections form the original omission set, R, which remains fixed during revision. The selector receives neither reference answers nor benchmark criteria.

You ask whether the selector can miss useful information or make incorrect selections, just as the model can make mistakes during initial writing. Writing requires the model to choose and organize information into an answer. The selector has a different task: it compares the completed report with the acquired findings to identify relevant information that is not fully expressed.

We considered these possible errors when designing the revision step. EGRC therefore treats the selections as candidate omissions, not as instructions to insert every item. During revision, the writer must identify a finding's existing expression, propose an edit, or explain why source, relevance, or space constraints prevent inclusion. Proposed edits undergo source review before they are applied.

Furthermore, we do not assume that the omission list is complete or error-free. To assess the selector's reliability, two independent human annotators checked a sample of 128 findings, as reported in the original submission (Section 3.2 and Appendix G). Both annotators confirmed source support, relevance, and absence for 40 of 64 selected findings, compared with 4 of 64 unselected findings. In this sample, the selected set thus concentrates human-confirmed omissions at ten times the rate of the unselected set, supporting the selector's ability to identify supported, relevant omissions.

Appendix E of the original submission also checks recovery using only omissions identified by both Qwen and DeepSeek. The two selectors agree on 57 omissions across 23 test tasks. We keep each method's reports unchanged and recalculate recovery using these findings. Recovery is 71.6% for EGRC, 31.2% for Generic, and 49.1% for Two-pass. The advantage therefore remains when both selectors must agree on the omissions being evaluated.

Section 3.1 of the revised manuscript now explains the selector's inputs, instructions, and outputs beside the definition of R. We have also added the original prompt and an input/output example to the supplementary material. The human comparison demonstrates the selector's ability to identify supported, relevant omissions. The two-selector check shows that EGRC's recovery advantage persists when evaluation uses only omissions identified by both Qwen and DeepSeek.

We have also clarified the revision procedure in Section 4, including how the writing LLM organizes the answer and decides what to do with each original omission. We now explain the separate LLM calls for checking edits against the sources and the program checks performed before edits are applied. These changes make the roles of the LLMs and the program explicit in the method description.

AuthorsOpenReview ↗

Responses to Reviewer tCrN (3/6)

3. Expression judgment M (W2; Methodology question 2).

Thank you for raising this question. M determines whether an omitted finding is counted as recovered, so its reliability is central to our evaluation.

For a text t and a finding e, an evaluation LLM determines M(t,e): 1 if the text fully expresses the finding, and 0 otherwise. In the main experiments, we use GLM-5.2 as the evaluation LLM. It receives the complete report, the findings being assessed, and their supporting source passages. It uses an evaluation prompt separate from the revision prompt. A positive judgment must quote text from the report that expresses the finding. A finding is counted as recovered only if the report itself fully expresses it. The writing LLM's own decisions do not determine the score.

The evaluation LLM checks the details essential to each finding's meaning. In the policy example in Figure 1, both the employee threshold and the subsidiary exclusion are needed to determine which firms qualify. A report stating only the threshold is therefore incomplete. The report can express these details in different words or across more than one passage. For example, the threshold can appear in one paragraph and the exclusion in another.

We validate these judgments against human annotations and a second assessment model. On 129 sampled report–finding pairs with valid model judgments, the evaluation LLM agrees with the two independent readers in 96.1% and 94.6% of cases, respectively (Appendix G.1). The readers also confirm a 34.4-percentage-point recovery advantage over Two-pass on the sampled DRB2 findings.

We also reassess the existing reports with GPT-5.5, comparing EGRC with List-only. Both receive the same organized answer and omission list, but List-only chooses which entries to address. Under this second assessment, EGRC's recovery is 17.3 percentage points higher across all 36 test tasks.

The original submission defined M in Section 3.1 and included the evaluation prompt in the supplementary package. We have expanded Section 3.1 to explain the evaluation inputs and judgment requirements more directly. Section 5.5 reports the human agreement, and Appendix C.2 gives the detailed expression criteria. Together, these results support the reliability of M and show that EGRC's recovery advantage is also observed in human judgments and with a different evaluation LLM.

Reviewer_jnE2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses a specific issue within Deep Research Agents: instances where an agent has successfully retrieved correct and relevant evidence, yet the final research report fails to fully incorporate it. The authors term this phenomenon "Evidence-to-Answer Attrition." To mitigate this, they propose a method called "Evidence-Grounded Residual Completion" (EGRC). This approach begins by structuring an answer based on the user's request; it then identifies findings omitted from the initial report and treats them as persistent constraints. For each omitted item, the model is tasked with making a "retain," "revise," or "defer" decision. Finally, the report is refined through source verification and targeted local edits. The method was primarily evaluated using the DRB2 benchmark.

Critical strengths
  1. The problem definition is of practical significance, and a series of analytical experiments confirm that this issue indeed exists within the current model.
  2. Extensive experiments conducted on datasets such as DRB2, scholarly synthesis, and SciConBench demonstrate the effectiveness of EGRC compared to other methods.
Critical weaknesses
  1. The scale of the test dataset is quite small. DRB2 consists of 132 samples. However, after the authors’ own partitioning into validation and test sets, followed by the filtering of specific omitted items, only 36 entries remain in the test set. While the reported performance gains may partly stem from this data filtering, the small sample size also introduces a high risk of volatility. The authors should consider reporting results based on the full DRB2 dataset. Furthermore, for the primary experimental results, it would be beneficial to include additional benchmarks, such as DRB1 and ResearchRubrics, rather than relying solely on DRB2. Incorporating a broader range of test sets would help strengthen the findings.
  2. EGRC is a method for further refining and modifying a model’s initial response; consequently, the gains it provides depend on the quality of that initial response. Since the current initial responses appear to be generated directly by a base model, a question arises: would these gains still be observed if the initial responses were instead generated by a "planning model" or a "Deep Research" model? This paper focuses on optimizing the final output generation stage of a Deep Research system. Therefore, demonstrating compatibility with existing Deep Research frameworks, such as WebWeaver,DeepWeaver or other method, would make the findings significantly more compelling.
  3. Most of the main experiments are based on Qwen3.8-Max, including the extraction of findings, identification of omissions, and generation of initial reports. On the one hand, this may introduce some bias into the overall pipeline. On the other hand, the authors should consider testing more models to demonstrate the transferability of the method, such as mainstream models from the GPT or Gemini series.
Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer jnE2 (1/4)

Dear Reviewer jnE2,

Thank you for recognizing the practical significance of evidence-to-answer attrition and the effectiveness of EGRC in our experiments. We appreciate your specific suggestions about task coverage, the quality of the starting report, and transfer across models. In response, we have clarified the task selection, examined the task-level gains, and completed additional experiments on DRB1 and on reports generated by DeepWeaver. We address your three main concerns in order below.

Responses to weaknesses and questions

Response to W1: test-set size, task selection, and evaluation on other benchmarks (and related questions).

Thank you for highlighting task coverage and the reliability of results from a small test set.

1. Why the test set has 36 tasks.

We use all 132 DRB2 tasks to study how often reports omit findings that the agent has already collected. To evaluate EGRC, we use a separate test set that was not used to develop the full method. We evaluate revision on reports where the selector finds omissions, because EGRC leaves the report unchanged when it finds none.

The tasks are divided as follows:

CollectionNumber of tasksUse in the study
Initial method design24Developing the initial method; not used for validation or test
Validation pool60Contains 44 tasks with detected omissions, used to validate the full method
Held-out test pool48Contains 38 tasks with detected omissions; two were blocked by provider input checks, leaving all 36 available tasks for evaluation

We fixed the held-out task pool and method configuration before running the comparisons. We did not select tasks based on any method's results. All methods are compared on these same 36 tasks. Among the 48 tasks reserved for testing, the selector found omissions in 38. Two of these tasks were blocked by the provider's input content moderation, so we evaluated all 36 remaining tasks.

EGRC revises a report only when the selector identifies omissions. Otherwise, it returns the original report unchanged. Across all 132 reports, the selector finds omissions in 106 and none in the other 26. Of these 26 reports, 16 are from the validation pool and 10 are from the test pool. They are not part of the 36-task comparison. We checked all 26 reports and confirmed that EGRC returned the original text without changes. Section 5.1 of the revised manuscript now explains which tasks are evaluated, and Appendix C.1 gives the task numbers above.

We agree that evaluation on more tasks and benchmarks is useful. We report the 44 validation tasks and 36 test tasks separately because they serve different purposes. A single score for all 132 tasks would mix tasks used during method development with tasks reserved for testing. Our analysis of omissions still covers all 132 tasks. The original submission also includes an evaluation on scholarly synthesis (Section 5.4). We have now added an EGRC comparison on 40 DRB1 tasks to evaluate the method on another benchmark for long research reports. We present these results below.

2. Do a few tasks account for most of the improvement?

Sections 5.2 and 5.3 of the original submission report the comparisons with Generic, Two-pass, and List-only. Generic chooses useful additions without receiving an omission list. Two-pass first asks the writer to find omissions and then add the missing content. List-only receives the same organized answer and omission list as EGRC, but chooses which entries to address. The table below summarizes the reported quality gains on the 36 test tasks:

EGRC compared withMean quality gain
Generic+4.7
Two-pass+5.0
List-only+3.9

All three quality gains are statistically significant in the task-paired analysis reported in Appendix E of the original submission (e.g., List-only: +3.9, 95% interval [2.7, 5.2]). The paper also reports that EGRC scores higher on 30 of 36 tasks against Generic, 32 against Two-pass, and 32 against List-only.

For this response, we also recalculated the average after excluding the three tasks with the largest quality gains over List-only. The average gain on the remaining 33 tasks was still above 3.1 points. We also checked the two languages separately, and the average gain is positive for both Chinese and English tasks. Thus, a few unusually large improvements do not explain the overall gain. These additional analyses use the scores from the original experiments. We have added them to Appendix E.

AuthorsOpenReview ↗

Responses to Reviewer jnE2 (2/4)

3. Experiments on DRB1 and scholarly synthesis.

Following your suggestion, we have evaluated EGRC on 40 DRB1 tasks, with 20 tasks from each language. We use findings, sources, and reports saved from the DRB1 agent before its final synthesis and rewriting. We compare the initial report with List-only and EGRC. The two revision methods receive the same organized answer, findings, and sources, and use the same revision procedure. EGRC treats every original omission as an explicit revision requirement; List-only can choose which omissions to address.

We evaluate quality with GPT-5.5 using the DRB1 RACE protocol. We compare each report with the benchmark reference in both presentation orders, then average the two scores. Reference reports and benchmark criteria are used only for evaluation. We assess recovery separately, using the same original omissions for all three methods.

DRB1 reportMean qualityRecovery (%)
Initial report52.90.7
List-only53.759.5
EGRC55.278.0

EGRC improves quality by 2.3 points over the initial reports. Compared with List-only, it improves quality by 1.5 points and recovery by 18.5 percentage points. Their average final report lengths differ by only 0.2%. EGRC scores higher in all four RACE dimensions, including content comprehensiveness and insight. These results show improvements in both quality and recovery on another benchmark, with the same starting answer and revision procedure for EGRC and List-only. We have added a summary in Section 5.6 and the setup and detailed results in Appendix I.

The original submission also evaluates 23 scholarly synthesis tasks constructed following DeepScholar-Bench, with a 600-word limit. The autonomous revision method starts from the same organized answer but selects its own findings. Compared with this method, EGRC improves recovery by at least 33 percentage points and achieves a statistically significant 8.8-point gain in reference-content coverage at nearly equal length. The content score measures coverage of key points from the reference related-work section, which the writing model does not receive. EGRC also receives higher content scores when the judge reads the report without the task abstract (Section 5.4). The scholarly experiment evaluates a different type of writing, while the new DRB1 experiment adds another benchmark for long reports. EGRC improves recovery and content quality in both settings. These results support its use across different Deep Research tasks, beyond DRB2. We agree that ResearchRubrics, which you also suggested, would further broaden the evaluation and leave it for future work.

AuthorsOpenReview ↗

Responses to Reviewer jnE2 (3/4)

Response to W2: starting-report quality and compatibility with DeepWeaver (and related questions).

We first clarify how the initial reports in our main experiments are generated. The system creates a section plan from the request and collects findings through search and source reading. A general-purpose large language model (LLM) writes each section using the findings and source information provided for that section. The completed sections are then combined into the initial report. Thus, these reports already use both planning and retrieved evidence. We have clarified this process in Section 5.1 and Appendix C.1 of the revised manuscript. On the test set, EGRC improves quality by 5.9 points over these initial reports.

We also checked whether the gains occur only on poor initial reports. We divided the test tasks into two equal groups based on their initial quality scores, without using the revision results. In the higher-scoring group of 18 tasks, EGRC raises average quality from 49.4 to 54.5, a gain of 5.1 points, and improves quality on 16 tasks. The lower-scoring group gains 6.7 points. Thus, EGRC also improves the better initial reports in the original experiment. We have added this analysis to Appendix E.

You also suggested evaluating EGRC on reports generated by existing Deep Research frameworks. We appreciate this suggestion, as it tests whether EGRC remains effective when another system generates the initial reports. We therefore applied EGRC to DeepWeaver's reports to test whether it could further improve both recovery and quality. We have added a summary in Section 5.3 and the setup and full results in Appendix E of the revised manuscript.

We use the DeepWeaver reports saved from our reproduction on the same 36 DRB2 tasks, with the acquired sources fixed. We run the selector on these reports to find their omissions. GPT-5.5 then organizes the answer and writes both revisions. EGRC and List-only receive the same organized answer, omission list, complete finding summaries, and sources, and use the same revision procedure. We also evaluate the unchanged DeepWeaver reports. This lets us compare EGRC both with the original DeepWeaver report and with List-only after the shared organization step.

We reuse the original quality assessment for the unchanged DeepWeaver reports. Recovery in all three conditions is evaluated on the same 630 omissions identified in these reports. The DeepWeaver reports have higher mean quality than the initial reports we generated for the main experiments (39.5 versus 36.7). The results are:

Report on the same 36 tasksMean qualityRecovery (%)Mean report length (tokens)
DeepWeaver initial report39.50.09,604
List-only revision41.040.712,578
EGRC revision43.361.412,749

EGRC improves the quality of the DeepWeaver reports by 3.9 points. When both revision methods start from the same organized answer and receive the same omission list, EGRC improves quality by 2.3 points and recovery by 20.7 percentage points over List-only. Both information recall and analysis improve, while average final lengths differ by only 1.4%. Quality gains are calculated before rounding the scores.

Because GPT-5.5 writes both revisions, the List-only comparison isolates the finding-level effect from the change of writer.

These results show that EGRC can further improve reports produced by an existing Deep Research framework. The gain over List-only shows the value of finding-level requirements beyond answer organization and access to the omission list.

Reviewer_4CrtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper, Found but Not Written: Understanding and Reducing Evidence Attrition in Deep Research, 1) proposes a new metric (recovery) to measure how much a generated deep research (DR) report fully expresses the key content given retrieved evidences; 2) proposes a new approach for a DP writer to revise its original report given explicitly identified omissions (i.e., content that is in a retrieved evidence, is relevant to the task, but is not fully expressed; 3) conducts experiments that shows that the proposed approach achieves the highest recovery score and report quality score comparing to alternative methods, based on a few DR benchmarks.

Critical strengths

The authors identify an important gap that is often missed in deep research evaluation - that is, often times a report might cover the correct retrieved evidences in general, but fails to include key details (or "completely expressed"). The study proposes both a metric to measure this aspect, and an approach to make the report writing better. The study shows thorough experiments comparing its proposed method with alternatives on multiple benchmarks.

Critical weaknesses
  1. While the big picture of the study is relatively well written, I find it difficult to understand the draft in many places, including some key parts. For example:
  • Section 4 is very hard to follow - many symbols introduced, some not well explained; terminologies mentioned without explanation (e.g. not sure what "plan membership" mean); unclear to me what the retain/revise/defer decisions mean or how do they alter the writer behavior, which is cited as one of the key part of the proposed method. Since Section 4 describes the key algorithm of the proposed report writing approach, I'd recommend significant rewriting before publication.
  • Section 3.3 requires significant effort to understand, and it's still not clear to me what "an equally sized, disjoint same-task finding set" actually look like.
  • Some more minor places that are hard to follow: in Section 1 (introduction), "target" is mentioned without explaining what it means (although it is explained later in the draft); Section 3.1 - unclear how R and M is calculated (it is later mentioned briefly that different LLMs are used, but unclear how they are prompted).
  1. While the paper considers the effect on the new report writing approach on report length, it doesn't show whether/how the report writing cost and latency are affected. I'd recommend these aspects be documented and discussed, as these are important for many DR applications.
Questions for authors

Main questions are to clarify the parts that I find hard to follow in the Weaknesses session above

6 comments
AuthorsOpenReview ↗

Responses to Reviewer 4Crt (1/5)

Dear Reviewer 4Crt,

Thank you for recognizing the importance of the problem and the range of experiments. We have revised the passages you identified to explain how the algorithm works, how we exchange the target lists, and how we measure omissions and their recovery. We have also added an analysis of inference cost and latency. Below we respond to each point and explain the changes to the manuscript.

Responses to weaknesses and questions

1. How the method works and what retain, revise, and defer mean (W1; Section 4).

Thank you for identifying the passages in Section 4 that need clearer explanations.

EGRC starts from an existing research report and the evidence already acquired for it. A finding is a specific statement drawn from that evidence, together with the conditions needed to interpret it. The aim is to recover relevant findings that the report has not fully expressed, without running another search.

Before revision, we prompt a large language model (LLM) to act as the selector. It reads the request, initial report, and summaries of the findings already collected, then selects findings that appear to be missing from the report. We keep this original omission list fixed during revision. The revision then follows the three steps shown in Figure 3.

First, organize the answer (Figure 3a). We prompt the writing model to identify parts of the user's request that the report has not adequately answered. It proposes an answer unit for each such part. In the illustrative policy example used throughout the paper (Figure 1), one such unit would explain which firms qualify for the policy. We then ask the model to align these units with the original omissions. The writer still has access to summaries of all acquired findings, including those already expressed in the report. It links the units to supporting findings and sources, reads the associated source passages, and proposes additions and corrections. Applying the accepted changes produces the intermediate report, denoted by y¹.

Regarding “plan membership” mentioned in your review, the original Section 4.2 used the policy example to explain why including a finding in the plan does not ensure its complete expression. Here, membership means that the plan links a finding to an answer unit. The eligibility unit may be linked to a finding containing both the employee threshold and the subsidiary exclusion, while the intermediate report states only the threshold. This is why EGRC revisits the original omissions after organizing the answer.

Second, decide how to handle every original omission (Figure 3b). We give the writing model the intermediate report and the original omission list. The revision prompt requires a decision for every entry in that list, returned together with the proposed local edits:

  • Retain: the intermediate report already conveys the finding. The model explains where it is expressed, so no additional text is needed for that finding.
  • Revise: the finding is still incomplete, so the model proposes local edits to complete it. In the same policy example, the proposed edit would add the missing subsidiary exclusion beside the employee threshold already stated in the intermediate report.
  • Defer: the finding is not added because the sources do not support its inclusion, it does not help answer the request, or it cannot fit within the length limit. The model must give a reason; it cannot simply skip an entry because it prefers another topic.

Because every entry requires an account, the writer cannot silently drop an omission. One edit can complete several related findings, and several passages can jointly express one finding. The requirement is to account for every original entry, not to append a separate sentence for each one.

Third, review and apply the proposed edits (Figure 3c). The program checks that every original entry has exactly one decision and that each revise decision refers to actual proposed edits. A separate LLM call reviews the edits against the source passages, checking support, scope, and preservation of useful existing content. Only edits that pass source review and fit the report-length limit are applied.

In the revised manuscript, Sections 4.1–4.3 describe these operations in order, explain what we ask the LLM to do at each step and which checks the program performs, and define symbols when they are introduced. Section 4.2 explains what each decision requires the writer to return and replaces “plan membership” with a direct explanation beside the policy example.

AuthorsOpenReview ↗

Responses to Reviewer 4Crt (2/5)

2. What is changed in the target-exchange experiment (W1; Section 3.3).

Section 3 studies how acquired findings are left out of reports. In these experiments, the writer adds content but does not change the original text. We exchange the target lists to test whether asking for different findings changes what the writer includes.

For each task, the original target list contains all omissions identified by the selector. We construct a second list of the same size using different findings from the evidence collected for that task. Where possible, we match each replacement to the same part of the request and report section as the original finding. No finding appears in both lists. For example, if the original list has two findings, e₁ and e₂, the replacement list contains two other findings, e₃ and e₄. “Same task” means that these all come from the evidence collected for the same research request, not from another question.

We compare two completion conditions using the same writing model. In the example above, one run is asked to add content for {e₁, e₂}, and the other for {e₃, e₄}. Each run receives the same initial report and all acquired finding summaries, including the findings in both lists. The writing budget and source-review rules are also unchanged.

For each output, we check both sets of findings: the original omissions and the replacement findings. For each set, we calculate the percentage fully expressed in the added text. We do not count the initial report in this assessment. The two conditions from Table 6 are shown below. Each row shows which list the writer received. The two result columns show how much of each set was fully expressed.

Findings explicitly requestedOriginal omissions fully expressed in additionsReplacement findings fully expressed in additions
Original omission list29.7%0.1%
Replacement list0.0%24.5%

We calculate the percentages for each task and then average over all 132 tasks. When we change the list, the writer includes more findings from the new list, although it still has access to the same evidence. This shows that the findings named in the list affect what is written. Section 3.3 now explains why we ran the experiment, how the two lists are formed, what stays the same, and what we evaluate.

3. What “target” means in the introduction (W1).

A target is a finding that we explicitly ask the writer to address. In targeted completion, we give the writer the omission list and ask it to add the relevant content. In untargeted completion, the writer receives the same complete finding summaries but chooses what to add without that list. Neither method searches for new evidence. We now explain “target” when it first appears in the introduction and describe the two methods before discussing their results.

AuthorsOpenReview ↗

Responses to Reviewer 4Crt (3/5)

4. How the omission set R and expression judgment M are obtained (W1; Section 3.1).

To obtain R, we ask an LLM to compare the findings with the initial report. The selector receives the user's request, the complete initial report, and summaries of all acquired findings, with their identifiers and source links. The prompt asks which findings contain useful information that the report has not yet expressed, including missing facts, corrections, quantities, comparisons, and qualifications needed to answer the request. The model returns finding identifiers, not a rewritten report. The program checks that every identifier is in the supplied collection and appears only once. The selector does not receive reference answers or benchmark scoring criteria.

M checks whether the generated text fully expresses a finding. For whole-report evaluation, we make a separate LLM call with the complete report, a finding, and its source passage. The model checks whether the report includes all essential details of the finding. A positive judgment must quote the relevant text from the report. The writing LLM's retain or revise decision does not determine this assessment. The source passage helps explain the finding, but information that appears only in the source does not count as being in the report.

When the evaluator gives a valid judgment, M = 1 means that the finding is fully expressed and M = 0 means that it is not. The completion experiments in Section 3 assess each addition on its own. The full revision experiments assess the complete final report.

The assessment checks whether the report preserves the essential facts and conditions stated in each finding. For a finding specifying an employee threshold and an exclusion, stating the threshold alone is incomplete because it changes which firms qualify. The details required for complete expression depend on the finding being assessed. Paraphrases are allowed, and several report passages can jointly express one finding. For example, an existing sentence may give the threshold and a later sentence may give the exclusion.

We check both procedures against human judgments. Readers check whether the selected findings are supported by sources, relevant to the request, and missing from the report (Section 3.2). They also judge whether the final reports express the findings (Section 5.5 and Appendix G). Section 3.1 now explains the inputs, instructions, and outputs for R and M. We have added the original selector prompt and an input/output example to the supplementary material. The prompt for assessing expression was already included with the original submission.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response and summary of revisions

Dear Reviewers and Area Chair,

Thank you for the reviews and discussion. We summarize the main clarifications and additional evidence incorporated in the revised manuscript.

Contribution and method. We study evidence-to-answer attrition: a Deep Research agent can find relevant evidence and still leave out details readers need. When the writer can only add text to the initial report, an omission list helps it fully express more missing findings, but mean report quality is no higher. This motivates Evidence-Grounded Residual Completion (EGRC). It organizes the answer around the request, then requires the writer to revisit each original omission: identify where it is already expressed, propose a local edit, or explain why it should be deferred. Proposed edits are checked against the sources. The revised introduction and Sections 3–4 define key terms, explain how omissions are identified and assessed, and distinguish the models' judgments from the program's checks.

On the 36 DeepResearch-Bench II (DRB2) test tasks, EGRC and List-only start from the same organized answer and share the omission list, evidence, writer, and revision procedure. List-only can choose which omissions to address. EGRC gains 3.9 quality points and 20.2 percentage points in recovery (the share of original omissions fully expressed in the final report). Mean final lengths differ by only 0.7%. Counting shared omission selection and answer organization, EGRC uses approximately 2.5% more total input and output tokens for the complete revision procedure.

Human comparisons also confirm that selected findings are more often genuine omissions than unselected findings. On 129 sampled report–finding pairs with valid model judgments, the expression assessor agrees with the two readers in 96.1% and 94.6% of cases, respectively (Section 5.5 and Appendix G). Report quality is assessed separately, without showing assessors the omission list.

Broader evaluation and transfer. The DRB2 test comparison includes all available held-out tasks with detected omissions (Section 5.1). The mean quality gain over List-only remains above 3.1 points after excluding the three tasks with the largest gains (Appendix E).

On 40 DeepResearch-Bench (DRB1) tasks, EGRC gains 1.5 points in mean quality and 18.5 percentage points in recovery over List-only, extending the evidence to a second long-report benchmark. On the 36 DRB2 test tasks, EGRC improves the mean quality of DeepWeaver's reports, showing that it can complement an existing Deep Research framework. It also gains 2.3 quality points and 20.7 percentage points in recovery over List-only, with GPT-5.5 writing both revisions. EGRC and List-only produce reports of nearly equal length in both settings. The GPT-5.5 and GLM-5.2 experiments show that the gains are not limited to the main Qwen writer (Sections 5.3 and 5.6; Appendices E and I).

Inference cost and practical value. Section 5.3 and Appendices E–F distinguish the cost of the complete EGRC revision from the additional cost of finding-level decisions, and report cumulative API time. We also directly compare Qwen-EGRC with the stronger GPT-6 Astra running lightweight Two-pass on 20 DRB2 test tasks. Qwen-EGRC gains 3.6 quality points for an additional estimated $0.28 per task before cache discounts, about 21% higher cost. These are public-price estimates from recorded token usage, including calls to auxiliary models and reasoning tokens; the accounting scope and unpriced failed requests are documented in Appendix F.

Update (Nov 19): In the follow-up discussion, Reviewer tCrN confirmed that the new comparison with GPT-6 Astra using Two-pass fully addressed their remaining concern about inference cost.

The central finding is that explicit decisions about each original omission improve both recovery and independently assessed report quality, even with the organized answer and omission list held fixed. This gives Deep Research systems a way to improve reports using evidence they already have, without another search. Thank you for considering the revised work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.