acceptodds
Under review as a conference paper at ICLR 2027

Don’t Waste Your Search-and-Visit Signals in Deep Research

Abstract

Deep research agents repeatedly search and visit the web, yet task-relevant signals exposed or derived during these interactions may not be retained in the updated research state. Representative approaches improve planning, exploration, or evidence organization, while systematic reuse of such unretained signals remains underexplored. We therefore introduce **WasteHub**, a plug-and-play mechanism that stores Unvisited-result, Visit-content, and Source-appraisal waste as typed, provenance-preserving residuals and selectively reuses them through a stage-aware gate without modifying the underlying framework's planner or search/visit interfaces. With ReAct as the underlying framework, WasteHub outperforms the Base setting and two baselines that augment Base with process reward model feedback (PRM-F) and post-answer exploration (PAE), respectively, on the aggregate metrics of DeepResearch Bench and DeepResearch Bench II. WasteHub uses fewer cumulative input tokens and has lower end-to-end latency than both PRM-F and PAE. Ablations show that all three residual types and both reuse phases contribute to the full method. Cross-framework experiments with Self-Manager and WebWeaver show that WasteHub achieves the best aggregate performance on both benchmarks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Deep Research, Large Language Model, Web Agents

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

This paper presents wastehub, a search info reuse framework that allows deep research agent to save and reuse useful search results. instead of discarding web searches and webpage visits, it stores three types of information: unvisited search results, omitted webpage content, and source-appraisal information. The stored information is selectively reused during later research steps to improve report quality and token efficiency. Experiments on deep research benchmarks (DRB and DRB II) show improvements in report quality and reduced resource consumption compared with baselines.

Critical strengths
  • Clear motivation for practical benefits. This is a real issue for deep research agents, where most search information is discarded during search and summarization.

  • Reasonable design with general applicability. The framework is a modular and general design and evaluated across multiple agent frameworks and models.

Critical weaknesses
  • The baseline selection is questionable. Evaluation mainly compares with PRM-F and PAE, which introduce additional feedback or exploration overhead. It is not clear why these two are most related direct comparison, instead of existing persistent-memory or retrieval-based mechanisms. Moreover, WasteHub does not show token savings over basic ReAct agent. Comparisons with simple memory-based methods would strengthen the work.

  • The performance improvement seems marginal relative to the additional complexity. The report-quality improvements over PAE are only 0.0089 on DRB and 0.0161 on DRB-II. Although statistically significant, their practical importance remains unclear. It is questionable whether these small improvements justify the extra system complexity. Also, does the framework gain applies to frontier LLMs like GPT5.6 or GPT6?

  • Unclear attribution of latency improvements. The evaluation shows end-to-end latency reductions, but the system relies on external search and webpage retrieval services. The authors use the same backend configurations, but authors do not clearly explain how variability of external service latency is controlled. This makes it difficult to attribute latency improvements specifically to WasteHub.

Questions for authors
  1. How does WasteHub differ from memory or retrieval-based mechanisms? Would a simpler memory cache achieve comparable results? Why are PRM-F and PAE the primary baselines? Given that WasteHub does not save tokens compared with ReAct, what is its practical advantage over the base agent?

  2. Can the authors provide evidence that the relatively small score improvements lead to clear differences in final report quality?

  3. How is external search and webpage retrieval latency variability controlled or measured? Can the authors report separate measurements for model inference, external tool calls, and WasteHub/framework-specific latency?

3 comments
AuthorsOpenReview ↗

Round1-Part1

We sincerely appreciate the time and effort you devoted to reviewing our paper, and your detailed, constructive suggestions.

Response to W1 & Q1: Baselines, distinctions, and practical benefits

Choice of PRM-F and PAE. WasteHub aims to improve report quality across deep research frameworks through a plug-and-play mechanism. We therefore compare enhancement strategies that can be naturally integrated into existing research workflows across frameworks. PRM-F improves research decisions through feedback, while PAE supplements evidence through additional exploration, representing two general enhancement strategies.

Distinction from memory and retrieval mechanisms. Memory mechanisms focus on storing, organizing, and updating information. For example, A-MEM organizes memory through note linking and updating. Retrieval mechanisms focus on recalling existing material relevant to current needs. WasteHub focuses on information omitted during search and visit: task-relevant information that is not retained in the native research state. It contributes a mechanism to preserve this information and reuse it at appropriate stages. Based on current needs, existing content, and information status, it selects residuals to guide verification, supplement evidence, or qualify claims, and plugs into different research frameworks.

Practical advantage over the base ReAct agent. WasteHub obtains larger quality gains with a smaller increase in input tokens. In the original main experiments, relative to Base, WasteHub increases cumulative input tokens by 17.0% / 17.1% on DRB / DRB-II and improves aggregate scores by 0.0148 / 0.0284. PAE increases input tokens by 67.9% / 43.2% and improves aggregate scores by 0.0059 / 0.0123. Thus, compared with additional exploration, reusing existing information improves the base agent's report quality with less additional input overhead.

Comparison with simpler memory. We have added A-MEM [1] and the Base Memory configuration from its original paper as direct controls. Base Memory retains note construction and retrieval but removes note linking and updates to existing memories, representing a simpler storage-and-retrieval approach. Both configurations use the public memory module adapted for use within a single ReAct task. Information is written after search and visit calls and retrieved before subsequent research and writing.

Table D1. Memory baseline comparison

Method DRB Overall DRB-II TotalScore
Base Memory0.47290.3126
A-MEM0.47610.3190
WasteHub0.48610.3380

WasteHub achieves higher mean scores on both metrics than the simpler Base Memory configuration (Table D1). With comparable total token usage, it also achieves higher mean scores on both metrics than A-MEM.

[1] Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-Mem: Agentic Memory for LLM Agents. In Advances in Neural Information Processing Systems (NeurIPS), volume 38, 2025.

AuthorsOpenReview ↗

Round1-Part2

Response to W2: Gains, overhead, and frontier models

Gains relative to added complexity. Score increments should be interpreted in terms of both the metrics and improvements in report content. DRB uses scores normalized relative to a reference report. With a fixed reference score, the same normalized gain requires a larger raw-score improvement at higher scores. On DRB-II, a gain of 0.0161 corresponds to an increase of 1.61 percentage points in the rubric pass rate averaged across tasks. At the content level, WasteHub improves Citation Accuracy from 48.78% to 63.95% and average Effective Citations from 6.67 to 8.19 over PAE (Table 4), reflecting improved source support in the reports.

In terms of resources, Table 1 and Table 2 show that, relative to Base, WasteHub achieves larger quality gains with a smaller increase in cumulative input tokens than PRM-F and PAE. At similar report quality, PAE-III uses about 2.01 / 1.85 times as many cumulative input tokens as WasteHub (Appendix H). These measurements include WasteHub's additional generation calls, showing that information reuse achieves report quality close to that obtained through additional exploration, with lower resource usage.

In terms of system design, WasteHub keeps the original planner and tool interfaces unchanged. The type ablations in Table 3 and stage ablations in Appendix G support the contributions of the information types and reuse stages. The added unified storage and retrieval control (Table D2) further shows that, with comparable total token usage, full WasteHub achieves higher mean scores on both metrics, supporting the additional quality benefits of its overall design over simple storage and retrieval.

Table D2. Unified storage and retrieval control

Setting DRB Overall DRB-II TotalScore
Unified storage and retrieval0.47400.3152
Full WasteHub0.48610.3380

Applicability to frontier backbones. Appendix D reports WasteHub's gains with Qwen3.5-397B-A17B, Kimi-K2.5, and DeepSeek-V4-Flash. To address your question, we have added experiments using GPT-6 Sol as the main agent, keeping the ReAct framework, tools, and auxiliary models unchanged.

Table D3. Results with GPT-6 Sol as the backbone

Method DRB Overall DRB-II TotalScore
Base0.53020.3554
PAE0.53510.3592
WasteHub0.54600.3731

With GPT-6 Sol, WasteHub achieves higher mean scores on both metrics than Base and PAE (Table D3).


Response to Q2: Improvements in report quality

Table 4 shows that, compared with PAE, WasteHub improves Citation Accuracy by 15.17 percentage points and provides an average of 1.52 additional source-supported citations per report, indicating improvements in both citation reliability and the number of effective citations.

The automatic score gains in the main experiments have statistical support and show a consistent direction across evaluator models. The main experiments use 5 independent runs, with task-level paired tests yielding p = 0.028 and p = 0.004 on the two benchmarks, respectively (Appendix E). When three evaluator models re-evaluate the same reports, WasteHub's gains over PAE remain within 0.0073 to 0.0104 and 0.0136 to 0.0194, respectively (Appendix L, Table 17).

We also conducted blind pairwise human evaluation by 3 independent raters, covering all 100 DRB tasks and 132 DRB-II tasks. We use each method's complete reports from its first run, hide method identities, and randomize presentation order. Raters compare factual coverage of the core question, source support, and analytical adequacy. For each preference judgment, raters must identify concrete content differences and verify the relevant original sources. They may choose a tie when they cannot clearly distinguish the reports. Results are determined by the majority judgment for each task, with a tie recorded when there is no majority.

Table D4. Blind human evaluation (task counts)

Benchmark Prefer WasteHub Prefer PAE Tie
DRB562420
DRB-II782925
Total1345345

Human raters prefer WasteHub on more tasks on both benchmarks (Table D4). These judgments, grounded in concrete content differences and source verification, together with the improvements in citation quality, support discernible quality improvements in the final reports.

AuthorsOpenReview ↗

Round1-Part3

Response to W3 & Q3: External-service variability and latency attribution

Web retrieval cannot fully eliminate variability from network conditions and real-time service load. To reduce its influence on method comparisons, we use the same search and page-fetching backends, concurrency settings, caching, and retry policies, keeping concurrency within the capacity of the experimental deployment. Methods are executed in randomized order within the same batch for each task, with 5 runs per method, to reduce the effects of service-load changes and execution order.

We also record the start and end times of each call. Table D5 breaks down latency into model calls from the original framework, external search and page fetching, and framework and WasteHub auxiliary processing. On both benchmarks, model calls account for approximately 80% of WasteHub's end-to-end latency.

Table D5. End-to-end latency breakdown (seconds per task)

Latency category DRB Base DRB WasteHub DRB-II Base DRB-II WasteHub
Original framework model calls1850.01710.42305.92298.7
External search and page fetching370.0365.0575.0580.0
Framework and WasteHub auxiliary processing25.7124.230.4158.7
End-to-end total2245.72199.62911.33037.4

On DRB, total model-call latency remains lower than Base even after including auxiliary calls, and its reduction is close to the reduction in end-to-end latency. Thus, the average latency improvement in these runs is primarily in model calls, while the combined time spent on external tools and non-model framework processing differs little.


Manuscript revisions

Considering all reviewers' feedback and the completeness and coherence of the manuscript, we have made the following revisions, all marked in blue:

  • Section 3.3: Clarified the rationale for baseline selection.
  • Section 5: Added discussion of memory-related baselines, such as A-MEM, and clarified their connections to and differences from our work.
  • Appendix G: Added the fixed-trajectory replay experiment with Synthesis residuals.
  • Appendix N: Added direct comparisons with task-conditioned summarization, ordinary page retrieval, and A-MEM.
Reviewer_Pa28OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work, the authors check how deep research agents are built around search and visit tools. It argues that a lot of useful information gets thrown away along the way. The authors define three kinds of "information waste". They propose WasteHub, a plug-in module that stores these as results with source provenance. A stage-aware gate filters them by type, alignment and novelty, and they are reinjected at five points: frontier cues for source selection, same-goal evidence reuse, appraisal notes on revisited sources, slot-matched synthesis before the report is written, and an appended source-scope paragraph. The planner and the tool interfaces of the underlying framework are not modified.The pipeline is evaluated on DeepResearch Bench and DRB-II. WasteHub improves over Base by about 0.015 on DRB Overall and 0.028 on DRB-II TotalScore. It also beats a PRM-feedback baseline and a post-answer exploration (PAE) baseline, and uses fewer tokens than either.

Critical strengths

The main strengths of the work are as follows:

I. In the field of deep research, this approach is interesting and quite novel. Using and revisiting sources in deep research pipeline has the potential to increase trust in the system.

II. The results are consistent were implementing this pipeline holds across three frameworks defined. Run-to-run std is around 0.002, and the paired test against PAE is significant on both benchmarks. However the absolute gains are small, so I would not trust a single-setting result. This consistency is the main reason I take the effect seriously.

III. The overall paper is well written with enough information in the appendix to help supplement the results.

Critical weaknesses

Given the strengths of the work, I still feel there are weaknesses that author needs to work on or address:

I. The baselines don't really test the main claim of the paper though. PRM-F and PAE are different kinds of improvement (feedback and more exploration). Neither is about reusing information the agent already retrieved. The comparisons I would expect are missing: (a) a less lossy visit step, for example summaries conditioned on the full task rather than only the current sub-goal, or a larger summary budget at matched tokens; (b) plain untyped retrieval over the visited pages with the same top-K; (c) some existing memory or workspace approach.

II. The results does not strongly tell whether the three-type taxonomy and the stage-specific gating matter, or whether a simpler "store more, retrieve later" design would perform about the same.

III. Report length and extra structure are not controlled for. WasteHub reports are about 50% longer than Base on DRB (18.7k vs 12.4k characters). RACE and rubric-recall judges are known to reward longer, more comprehensive outputs, and the biggest gains are in Comprehensiveness and Information Recall. So I am not sure what the main reason could be here.

IV. Just using DRB and DRB-II might not be enough as it only consistently checks the overall quality of the work. The parameters are still very open to interpretation with most of the work in RACE parameters being evaluated by just an AI judge. Maybe an in-house evaluation setup would help to just capture the difference that the paper wants to focus on as these improvement could mean anything.

V. The related work also leaves out the LLM agent memory literature (MemGPT, A-MEM, ReadAgent, etc.), which I think is the closest prior work and can be used in the pipeline.

Questions for authors

Please refer to the weaknesses mentioned. However these are some of the questions that I thought would need some clarifications:

I. In Table 2, WasteHub has lower end-to-end latency than Base on DRB (2199.6s vs 2245.7s), even though it uses more rounds and runs an extra extraction call over each visited page. Can you break down tokens and latency into main agent, extraction and gate calls?

II. How does WasteHub compare, at matched token cost, with (a) a visit summarizer conditioned on xx as well as gtg_t, and (b) untyped top-KK retrieval over visited page chunks? These results would have the biggest effect on my understanding of this work and its impact.

III. Can you report a length-controlled comparison, e.g. score regressed on report length, or WasteHub vs PAE-III at similar length? What is the score if the Boundary paragraph is removed before evaluation?

IV. Noise-WasteHub is slightly below Base, even though everything in its pool passed the task-relevance check at collection time. Does this mean most stored residuals are of little use? What fraction of collected residuals are ever selected?

4 comments

Thank you for your careful reading and thoughtful comments. Your feedback has been very helpful in improving the manuscript.

Response to W1 & Q2: Direct baselines and comparable token usage

WasteHub aims to improve report quality across deep research frameworks by reusing acquired information through a plug-and-play mechanism. We therefore compare enhancement strategies that can be naturally integrated into existing research workflows across frameworks: PRM-F improves research decisions through feedback, while PAE supplements evidence through additional exploration. These comparisons assess the quality gains and resource requirements of these enhancement strategies.

We have added three controls:

  • Task-conditioned summary: The visit summarizer is conditioned on both the full task description xx and the current sub-goal gtg_t.
  • Ordinary page retrieval: Untyped top-K retrieval over chunks of visited pages, using the same top-K and injected-token cap as the corresponding WasteHub stage.
  • A-MEM [1]: A within-task adaptation of the public memory module to ReAct, retaining note construction, linking, and updating. Information is written after search and visit calls and retrieved before subsequent research and writing.

All methods use the same model and tool configurations, with comparable total token usage.

Table P1. Direct baseline comparison

Method DRB Overall DRB-II TotalScore
Task-conditioned summary x+gtx+g_t0.47080.3089
Ordinary page top-K retrieval0.47320.3131
A-MEM (within-task)0.47610.3190
WasteHub0.48610.3380

These controls test full-task conditioning during summarization, retrieval from visited pages, and memory organization through linked notes, respectively. With comparable token usage, WasteHub achieves higher mean scores on both metrics (Table P1), supporting the value of selectively preserving and reusing omitted information.

[1] Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-Mem: Agentic Memory for LLM Agents. In Advances in Neural Information Processing Systems (NeurIPS), volume 38, 2025.


Response to W2: Information types and stage-aware gating

The three residual types serve different purposes: Unvisited-result residuals guide subsequent verification, Visit-content residuals supplement evidence, and Source-appraisal residuals constrain source reuse and the scope of claims. WasteHub therefore applies different collection and reuse rules to them. The type ablations in Table 3 and stage ablations in Appendix G show lower mean scores when the corresponding information type or reuse stage is removed.

We have also added a “store more, retrieve later” control. It stores search results and visited-page content together, then retrieves and injects information by relevance before each research decision and final writing, without distinguishing the three residual types or applying stage-aware gating. Both settings use the same underlying framework, model, and tool configurations, with comparable total token usage.

Table P2. Unified storage and retrieval control

Setting DRB Overall DRB-II TotalScore
Unified storage and retrieval0.47400.3152
Full WasteHub0.48610.3380

With comparable token usage, WasteHub achieves higher mean scores on both metrics than the unified storage and retrieval control (Table P2), supporting the value of organizing information by type and reuse stage.

Response to W3 & Q3: Report length and Boundary controls

Appendix H, Figure 5 already provides a comparison with PAE-III. WasteHub's reports contain 18.7k / 19.3k characters on DRB / DRB-II, versus 20.2k / 20.8k for PAE-III. Their mean scores are close, differing by 0.0007 / 0.0009. WasteHub thus achieves report quality close to that of PAE-III, which performs more extensive exploration, with slightly shorter reports.

Appendix G, Table 11 reports the ablation with Boundary disabled. This stage only appends a paragraph describing source applicability at the end of the report, without modifying the preceding body. Disabling it yields DRB Overall and DRB-II TotalScore of 0.4846 and 0.3351, decreases of 0.0015 / 0.0029 from the full configuration.

We further control report length and Boundary jointly. We retain each method's existing research materials and regenerate reports using a common writing template, removing WasteHub's Boundary paragraph before evaluation. Length matching is based on the report bodies actually evaluated for each method.

Table P3. Length matching and Boundary control

Method DRB Overall DRB-II TotalScore
Base0.47280.3115
PAE-I0.47960.3227
PAE-III0.48570.3359
WasteHub, Boundary removed0.48520.3366

With comparable body lengths and no Boundary paragraph, WasteHub still achieves higher mean scores on both metrics than Base and PAE-I, and remains close to PAE-III (Table P3). Separately, the original experiments in Appendix H show that PAE-III achieves report quality close to WasteHub while using more cumulative input tokens and taking approximately 1.78 and 2.19 times as long end to end on DRB and DRB-II, respectively.


Response to W4: Evaluation of report content

The two benchmarks themselves cover rich, complementary dimensions: Comprehensiveness, Insight, Instruction Following, and Readability, as well as Information Recall, Analysis, and Presentation (main-text Table 1), together with Citation Accuracy and Effective Citations (Appendix A). These metrics characterize differences in information coverage, analytical quality, and source support. Although some metrics are assessed by AI judges, all methods are compared using consistent, clearly specified scoring criteria and evaluation procedures (Appendix C). The main experiments include 5 independent runs and task-level paired tests (Appendix E), and gains persist across backbones (Appendix D). Re-evaluation by different evaluator models, with the same reports, scoring criteria, and weights held fixed, also yields a consistent direction of improvement (Appendix L).

We also conducted blind pairwise human evaluation by 3 independent raters, covering all 100 DRB tasks and 132 DRB-II tasks. We use each method's complete reports from its first run, hide method identities, and randomize presentation order to compare PAE and WasteHub on the same task. Raters assess factual coverage of the core question, source support, and analytical adequacy to determine which report provides a more complete and evidence-supported answer. For each preference judgment, raters must identify concrete content differences and verify the original sources supporting it.

Table P4. Blind human evaluation (task counts)

Benchmark Prefer WasteHub Prefer PAE Tie
DRB562420
DRB-II782925
Total1345345

Human raters prefer WasteHub on more tasks on both benchmarks (Table P4). These judgments, grounded in concrete content differences and source verification, agree with the direction of the automatic evaluation gains and provide independent human support for WasteHub's report-quality advantage.

Response to W5: Related work on agent memory

We have added MemGPT, A-MEM, and ReadAgent to Section 5 and clarified their connections to our work. They focus on hierarchical context management, dynamic linking and updating of memory notes, and long-document processing through compressed memories and rereading the original text, respectively.

These works address context management, memory organization, and long-document reading. Our work focuses on information present in tool returns but omitted from the native research state in deep research workflows: content that has already been acquired and is relevant to the task but has not been retained. WasteHub preserves this omitted information and reuses it according to its type and the subsequent research stage to guide verification, supplement evidence, or qualify claims. It keeps the original planner and tool interfaces unchanged and can be plugged into different research frameworks.

We select A-MEM as the representative memory baseline in Table P1 because it dynamically organizes information through note linking and memory evolution, and its modular implementation is relatively well aligned with our plug-and-play design across research frameworks.


Response to Q1: Token and latency breakdowns

Table P5 and Table P6 report total token usage and latency, respectively, grouped into model calls from the original framework, residual extraction, and gating and other auxiliary calls.

Table P5. Total model token usage breakdown (thousand tokens)

Call category DRB Base DRB WasteHub DRB-II Base DRB-II WasteHub
Original framework model calls419.2351.3522.8452.6
Residual extraction0116.90135.6
Gating and other auxiliary calls017.5019.6

Table P6. Model latency breakdown (seconds per task)

Call category DRB Base DRB WasteHub DRB-II Base DRB-II WasteHub
Original framework model calls1850.01710.42305.92298.7
Residual extraction066.8079.1
Gating and other auxiliary calls030.4048.5

On DRB, both total token usage and latency of the original model calls decrease. The 139.6-second reduction offsets the 97.2-second overhead of the added auxiliary calls. On DRB-II, the original model-call latency remains close to Base, while the added auxiliary calls increase total model-processing time.


Response to Q4: Random reuse and selection rates

Deep research proceeds through a sequence of sub-goals. Admission checks whether information is relevant to the full task, whereas reuse additionally requires assessing its relevance to the current sub-goal and whether it adds to existing content. Thus, residuals that pass task-relevance checks can still be off-target or redundant at a particular stage, and random injection need not improve the report. Noise-WasteHub's slightly lower performance than Base does not imply that most residuals lack value.

In the full WasteHub main experiments, after deduplicating record IDs within each run and then aggregating across runs, 26.2% / 23.5% of records on DRB / DRB-II were selected by a gate at least once. Full WasteHub improves report quality at these selection rates, whereas random reuse does not reproduce the gains. This indicates that matching candidate information to current research needs is important for realizing the benefits of residual reuse.


Manuscript revisions

Considering all reviewers' feedback and the completeness and coherence of the manuscript, we have made the following revisions, all marked in blue:

  • Section 3.3: Clarified the rationale for baseline selection.
  • Section 5: Added discussion of memory-related baselines, such as A-MEM, and clarified their connections to and differences from our work.
  • Appendix G: Added the fixed-trajectory replay experiment with Synthesis residuals.
  • Appendix N: Added direct comparisons with task-conditioned summarization, ordinary page retrieval, and A-MEM.

Dear Reviewer,

Thank you again for your thoughtful review. As the discussion period is nearing its end, we would like to kindly follow up on our rebuttal and ensure that we have adequately addressed your concerns.

We would greatly appreciate any further feedback you may have, and we would be happy to clarify any remaining questions.

Thank you for your time and consideration.

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

In a deep-research system, the agent searches for sources, visits selected pages, summarizes what it finds, and uses that working state to guide further research and write a report. Some information is left out of the active state because it is not needed for the current step, not necessarily because it is irrelevant. WasteHub preserves these details with their sources and selectively retrieves them later:

  • Before choosing a source, it can surface unvisited search results as leads. These still need to be checked before they can support a claim.
  • During later research, it can retrieve details omitted from page summaries when they match the current research goal.
  • During report synthesis, it can bring back omitted page details and notes about a source’s scope, such as its date, geographic coverage, or definitions, to help qualify claims.

Main contribution: a residual-memory mechanism for reusing task-relevant information that the underlying agent encountered but did not carry into its normal research state. It is designed to plug into existing search-and-visit frameworks - WasteHub uses Qwen3.5-35B-A3B for its auxiliary operations to select relevant residuals, then injects only those selected records into later contexts for the research agent.

Critical strengths
  • The work presents an interesting idea to treat information omitted from the agent’s active research state as potentially useful later, rather than as permanently irrelevant, and to preserve it for selective reuse.
  • The paper examines the method through ablations that remove residual types or restrict the stages where residuals can be reused. The results show how performance changes under these variants; the Goal and Synthesis stages have the largest reported score drops when disabled.
Critical weaknesses
  • Gains over the strongest baseline, PAE, are positive but modest on DRB, though larger on DRB-II. The real-world importance is fully established with the current experiments.
Questions for authors

Could you evaluate the value of retrieved residuals with a matched-trajectory or replay-based ablation, for example, holding the search and visit history fixed while withholding selected residuals from a later stage, to better isolate whether the information itself improves the report, separately from changes to the agent’s subsequent research behavior?

2 comments

Thank you for your encouraging assessment of our work and your valuable suggestions for improving the manuscript.

Response to W1: Gains and practical significance

The main experiments include 5 independent runs, and WasteHub's gains over PAE on both benchmarks are statistically significant under task-level paired tests (Table 1, Appendix E). Across frameworks, backbones, and evaluators, WasteHub also outperforms PRM-F and PAE on both aggregate metrics, showing a consistent direction of improvement across configurations.

The two benchmarks use different scoring scales, so their absolute gains are not directly comparable. DRB Overall is a quality score normalized relative to a reference report: a score of 0.5 on an individual task means that the generated and reference reports receive the same score under the same evaluation criteria. For a given task with a fixed reference score, achieving the same normalized score increment requires a larger raw-score improvement at higher scores. DRB-II TotalScore is the rubric pass rate averaged across tasks, where an increase of 0.01 corresponds to 1 percentage point (Appendix C). Thus, a smaller numerical gain on DRB than on DRB-II does not directly imply a smaller improvement in report quality.

Citation quality also improves. Compared with PAE, Citation Accuracy increases from 48.78% to 63.95%, and average Effective Citations increases from 6.67 to 8.19 (Table 4). Compared with PAE-III, which achieves similar report quality, WasteHub uses 441.9k / 550.2k cumulative input tokens, versus 887.9k / 1,015.8k for PAE-III (Appendix H), illustrating the benefits of information reuse in terms of quality and resource usage.


Response to Q1: Fixed-trajectory replay

Appendix G, Table 11 shows that removing only the Synthesis-selected residuals while retaining the report outline reduces DRB Overall and DRB-II TotalScore by 0.0051 and 0.0093, respectively, providing evidence for the contribution of information reuse during writing.

To isolate the contribution of these residuals to the report, we also conducted paired replay with fixed research trajectories. At the checkpoint after research ends but before Synthesis residuals are injected, we fix the research state and outline of each WasteHub trajectory and generate reports under the three conditions in Table R1. All three share exactly the same search and visit history and differ only in the residual information added for writing. They use identical writing settings, with Boundary disabled and no further search or visit calls.

The random and Synthesis conditions select content from the same set of stored Visit-content residual records. The random condition samples the same number of records as the Synthesis condition, with approximately the same total length of added text.

Table R1. Fixed-trajectory replay

Information added for writing DRB Overall DRB-II TotalScore
No Synthesis residuals0.48060.3274
Random Visit-content residuals0.47980.3290
Synthesis-selected residuals0.48490.3356

In this paired replay with fixed research histories, the Synthesis-selected condition achieves higher mean scores on both metrics than either the no-addition or random condition (Table R1). This supports the contribution of selected residuals to subsequent report generation from a fixed research state.


Manuscript revisions

Considering all reviewers' feedback and the completeness and coherence of the manuscript, we have made the following revisions, all marked in blue:

  • Section 3.3: Clarified the rationale for baseline selection.
  • Section 5: Added discussion of memory-related baselines, such as A-MEM, and clarified their connections to and differences from our work.
  • Appendix G: Added the fixed-trajectory replay experiment with Synthesis residuals.
  • Appendix N: Added direct comparisons with task-conditioned summarization, ordinary page retrieval, and A-MEM.

Dear Reviewer,

Thank you again for your thoughtful review. As the discussion period is nearing its end, we would like to kindly follow up on our rebuttal and ensure that we have adequately addressed your concerns.

We would greatly appreciate any further feedback you may have, and we would be happy to clarify any remaining questions.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.