acceptodds
Under review as a conference paper at ICLR 2027

WideResearcher: Knowledge-Graph-Guided Deep Research for Broader Exploration

Abstract

Deep research tasks require an AI system to synthesize comprehensive, insightful long-form reports through extensive web search across tens to hundreds of sources. Despite steady recent progress, current agentic deep research systems still suffer from two major problems. First, the initial batch of search queries which strongly shapes the entire research trajectory can only depend on the parametric knowledge of the Large Language Models (LLMs), leading agents into deep "rabbit holes" at the expense of topic coverage breadth. Second, relations between different parts of the knowledge landscape are not easily discoverable from unstructured text, making it difficult for agents to recognize what they do not know, and consequently, which topics to explore next. To tackle these problems, we propose WideResearcher, a hierarchical multi-agent deep research system capable of performing high-quality research broadly and deeply. Our system conducts a preliminary exploration of an external knowledge base before any web search takes place, relieving the hard dependency on the LLM's parametric knowledge and ensuring broad coverage of the topic's knowledge landscape. A shared knowledge graph is then maintained throughout the research process, condensing unstructured web articles into structured entities and relations with provenance groundings, whose semantic and topological properties are exploited to guide subsequent research directions. We evaluate WideResearcher against a range of proprietary and open-source deep research systems on two established deep research benchmark, LiveResearchBench and DeepResearch Bench. Experimental results show that our system consistently produces higher-quality reports compared to existing approaches, showing the effectiveness our knowledge-graph-guided approach to deep research. The source code of WideResearcher is provided as supplementary material and will be publicly released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Deep Research, Agent, Large Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_beJ9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors argue two limitations of existing work: (1) the initial batch of search queries can not be accurate because they can only depend on the parametric knowledge of LLMs; (2) relations between different knowledge landscapes are difficult to discover. So the authors propose wideresearcher, a hierarchical multi-agent deep research system, the system can perform high-quality research broadly and deeply. They conduct experiments on two established deep research benchmarks.

Critical strengths
  1. The question is clear. One is the initial query can only depend on the LLMs inner knowledge, the second one is the intermediate memories are unstructured texts.
  2. The writing is easy to follow, figure 1 can clearly illustrate the motivation
  3. The experimental results show a significant enhancement in benchmark LiveResearch.
Critical weaknesses
  1. In this paper, the authors claim that they try to solve the problems from both depth and breadth, but the broad way can not be verified from the main results. For example, wideresearcher is weaker than Webweaver on coverage.
  2. The ablation study is incomplete. It includes only two variants (w/o Seeding and w/o shared KG) and reports only the coverage metric on LiveResearchBench.
  3. The experimental section is relatively thin. Key analyses are missing, such as hyperparameter sensitivity experiments.
  4. Using knowledge graphs in deep research already seems to be an established sub-area. Why didn't the authors compare against more similar methods in the main expeirmental results, such as [1]?
  5. The claim "To our knowledge, we are the first to investigate deep research beyond depth alone" is too strong.
  6. The cost of wideresearcher is way higher than the other baselines. Are there any methods to lower the cost?

[1] Shi, Zhuofan, Ming Ma, Zekun Yao, Fangkai Yang, Jue Zhang, Dongge Han, Victor Rühle, Qingwei Lin, Saravan Rajmohan, and Dongmei Zhang. "A Tale of Two Graphs: Separating Knowledge Exploration from Outline Structure for Open-Ended Deep Research." arXiv preprint arXiv:2602.13830 (2026).

Questions for authors

please refer to the weakness section

2 comments

Thank you for your thoughtful review and constructive feedback. We respond to your questions below.

  1. On the evaluation of research diversity and breadth: The "Coverage" metric in LiveResearchBench and the "Comprehensiveness" metric in DeepResearchBench are the closest available proxies we could find for research diversity and breadth. However, because these metrics are inherently recall measurements over predefined lists curated by the benchmark authors, they may not fully capture or reward systems designed to favor greater breadth along the breadth-depth spectrum.
  2. On the issue of too few variants in the ablation study, we are conducting an additional ablation study on a variant of WideResearcher with the knowledge-gap detection module removed. We will report the results as soon as the experiments are complete.
  3. We reproduced DualGraph on LiveResearchBench using the same model, DeepSeek-v4-Flash-0731, as in our experiments. The official repository contained missing functions and was not directly runnable, we therefore made a best-effort attempt to complete the implementation. The reproduced DualGraph outperformed the other baseline systems reported in our paper, including GPT-Researcher, which achieved an average score of 67.0, the highest previously reported except from our system. Nevertheless, its performance still remained substantially below that of WideResearcher. We will include the complete reproduction details and results in the revised manuscript.
Agent SystemAveragePres.Cons.Cov.Cit.
DualGraph74.373.251.689.982.5
WideResearcher (Ours)82.189.966.987.583.9
  1. About the overly strong claim: We will revise it to better reflect the scope of our results. Thank you for pointing this out.
  2. On token cost and runtime efficiency: We acknowledge that the token cost of our system is substiantially higher than other systems, and we will include the cost-analysis in the main text in our revisions. The high token usage is mainly caused by our parallelized subagent design, where multiple agents concurrently modify the same shared knowledge graph, and each digester agent must access the latest version of the graph before expanding it to reduce extraneous insertions. Efficiently coordinating asynchronous updates to shared memory remains a nontrivial challenge and is beyond the scope of the current work. That being said, in our attempt to reproduce the results of DualGraph, we observed that DualGraph required more than 3 hours per query on average, compared with approximately 1.5 hours for our system under the same experimental settings. These observations suggest that, despite its higher token consumption, our system remains competitive in terms of wall-clock runtime. Nevertheless, we agree that reducing token usage is an important direction for future research.

Thank you for your response. I look forward to seeing the new results. For now, I am inclined to keep my score.

Reviewer_KpzwOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper identified two limitations of current deep research systems. First, initial research directions depend on the model's parametric knowledge, potentially limiting exploration breadth. Second, many existing systems use unstructured memory which makes it hard to track explored topics, identify knowledge gaps, and guide subsequent searches. To address these limitations, the paper proposes WideResearcher, which uses an external knowledge base to seed the initial exploration and use a knowledge graph to maintain research state and guide further searches through knowledge-gap discovery.

Critical strengths

The premise of using a knowledge graph to track the research process is sound. I particularly appreciate the discussion on knowledge graph discovery, where the system flags semantically related topic clusters never connected”, and “flags fragile connections warranting more searches.”

The paper’s writing and presentation is clear.

Critical weaknesses
  1. Limited empirical evidence for improved research breadth The empirical results do not support the paper's central claim of improving research breadth. On DeepResearch Bench, WideResearcher achieves a Comprehensiveness score of 54.4, slightly below Open Deep Research's 54.6. The improvement in RACE Overall is also marginal (53.9 vs. 53.2). On LiveResearchBench, WideResearcher achieves Coverage of 87.5, below WebWeaver (88.5) and only marginally above Open Deep Research (86.7). While the system achieves substantial improvements in other metrics, particularly citation quality, these do not directly establish the claimed improvement in research breadth. This is important because broader exploration is a primary motivation for the proposed method.

  2. Limited evidence from the ablation study: The ablation study provides limited evidence for the effectiveness of the proposed components in improving research breadth. On LiveResearchBench, removing the shared KG decreases Coverage from 87.5 to 85.3 (2.2 points), while removing external-KB seeding decreases it only to 87.2 (0.3 points). In particular, the small improvement from external-KB seeding raises questions about one of the paper's primary contributions. The additional Vendi score analysis shows only a modest increase in topical diversity (11.33 vs. 10.89). However, this metric measures diversity among cited article summaries rather than the relevance or completeness of the explored topics. It is therefore unclear whether this increase is due to higher coverage of relevant topics or could also include higher coverage of other, potentially irrelevant topics. Furthermore, the ablation study does not separately isolate the contribution of knowledge-gap discovery, making it difficult to assess the effectiveness of this mechanism independently of other benefits of the shared KG.

  3. Substantially higher inference compute: Table 3 appendix shows that WideResearcher consumes substantially more tokens compared to baselines. Despite the substantially higher compute budget, the method achieves little to no improvement in breadth-related metrics (point 1 above), further weakening the empirical support for its central claim. The token counts should be included in main text along with cost and latency metrics, to make the quality-efficiency tradeoff more transparent.

Questions for authors

deal with a lot of unstructured data and concepts all at once. However, it is likely that there are problems for which knowledge graphs are the best representation. Can the authors define the scope? For example, have you performed analysis of the benchmarks and identified the tasks for which your method excels? If so, perhaps the authors may want to focus on those problems for which this technique is applicable.

2 comments

Thank you for your thoughtful review and constructive feedback.

  1. On token cost and runtime efficiency: We acknowledge that the token cost of our system is substiantially higher than other systems, and we will include the cost-analysis in the main text in our revisions. The high token usage is mainly caused by our parallelized subagent design, where multiple agents concurrently modify the same shared knowledge graph, and each digester agent must access the latest version of the graph before expanding it to reduce extraneous insertions. Efficiently coordinating asynchronous updates to shared memory remains a nontrivial challenge and is beyond the scope of the current work. That being said, in our recent attempt to reproduce the results of the DualGraph deep research system, we observed that DualGraph required more than 3 hours per query on average, compared with approximately 1.5 hours for our system under the same experimental settings. These observations suggest that, despite its higher token consumption, our system remains competitive in terms of wall-clock runtime. Nevertheless, we agree that reducing token usage is an important direction for future research.
  2. On evaluation of research diversity and breadth: Our system was originally motivated by the goal of improving deep research agents for quantitative finance, where it is important to cover a broad knowledge landscape while reasoning over long causal chains. For example, a certain geopolitic conflict may cause the blockage of an important passage for oil transportation, which causes the gas price to increase, which in turn makes EV or hybrid vehicles a more economical choice. In this scenario, although the causal chain is highly indirect, it is still non-negligible thus needs to be covered in industry chain analysis. The "Coverage" metric in LiveResearchBench and the "Comprehensiveness" metric in DeepResearchBench are the closest available proxies we could find for research diversity and breadth. However, because these metrics are inherently recall measurements over predefined lists curated by the benchmark authors, they may not fully capture or reward systems designed to favor greater breadth along the breadth-depth spectrum. Although our system remains competitive on these metrics, we agree that they may not fully align with the specific capabilities our work aims to promote. We greatly appreciate your suggestions and will consider a clearer and more focused framing of our contributions.
  3. On effectiveness of knowledge-gap discovery: To assess the contribution of knowledge-gap discovery, we are conducting an additional ablation study on a variant of WideResearcher with the knowledge-gap detection module removed. We will report the results as soon as the experiments are complete.

Thanks for the response. I look forward to seeing your new results, and I'm keeping my original rating in the meantime.

Reviewer_BGx8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces WideResearcher, a hierarchical multi-agent system that uses knowledge graphs to improve deep research. The authors identify two limitations of existing approaches: reliance on LLM parametric knowledge for initial research planning and difficulty identifying knowledge gaps from unstructured research histories. WideResearcher first builds a seed knowledge graph from Wikidata, expands it with information extracted from web searches, and uses graph structure to identify important entities and potential knowledge gaps to guide further research. The authors evaluate WideResearcher on DeepResearchBench and LiveResearchBench. The system outperforms the reproduced open-source baselines in overall scores, with particularly strong gains in citation-related metrics. Ablation studies show modest improvements in coverage and topical diversity from the proposed components.

Critical strengths
  1. Well-motivated research problem. The paper identifies meaningful limitations of existing deep research systems, particularly their dependence on initial research directions and the difficulty of tracking relationships across large collections of retrieved information. Improving exploration breadth and evidence organization is relevant to practical deep research applications.
  2. WideResearcher combines Wikidata seeding, dynamic graph expansion, and knowledge-gap discovery to explore new research directions. Its provenance tracking also helps link findings to their original sources.
Critical weaknesses
  1. Limited evidence that knowledge-base seeding improves research breadth. The paper uses Wikidata to build the initial knowledge graph. However, Wikidata covers a wide range of topics, and not all retrieved entities and relationships may be relevant to the research question. The paper does not clearly evaluate how well the system filters irrelevant information while retaining useful connections. More importantly, removing Wikidata seeding only reduces coverage from 87.5 to 87.2 on LiveResearchBench. This small difference raises questions about how much seeding actually contributes to broader research.
  2. Lack of direct validation of knowledge-gap discovery. The system identifies knowledge gaps by looking for disconnected or weakly connected groups of entities in the graph. However, a missing connection does not necessarily mean that important information is missing. It could simply be caused by incomplete information extraction or graph construction. The paper does not evaluate whether these identified gaps are actually useful or whether exploring them helps discover relevant information that would otherwise be missed.
Questions for authors

How sensitive is WideResearcher’s performance to the quality and relevance of the initial Wikidata subgraph? How does the KB explorer determine which Wikidata entities and relationships are relevant to the research question? Does graph expansion amplify irrelevant connections?

1 comment

Thank you for your thoughtful review and constructive feedback.

  1. On Weakness 1 and Questions about relevance filtering during external KB exploration: In short, our system relies on the LLM to determine which SPARQL queries to issue and which entities and relations are relevant to the original research query. We made this design choice for the following reasons:

    1. We believe that the LLM’s understanding of natural language provides a flexible alternative to standalone semantic relevance methods, such as those based on vector embeddings, when assessing the relevance of entities and relations, because the relevance of a piece of information to a topic is highly contextual.
    2. The open-ended nature of deep research tasks makes it difficult to define a clear boundary for topic relevance. The same piece of information may be considered irrelevant for a shallow survey but useful for a deep dive. For such tasks, we therefore prioritize information recall across long-distance relationships, as overly strict filtering could discard information that becomes valuable later in the research process.

    For these reasons, we chose not to introduce additional explicit relevance filters into the external KB exploration step. Nevertheless, we agree that the trade-off between recall and precision is important, requiring further investigation.

  2. On Weakness 2 about reliability and effectiveness of knowledge-gap discovery:

    1. Regarding the concern that apparent knowledge gaps may result from incomplete information extraction or graph construction rather than genuinely missing information, both the digester and researcher agents have access to the same knowledge-gap detection tools. Before ending its session, a digester can invoke these tools to check whether any remaining gaps can be addressed using the article it is currently processing. This reduces the likelihood that "false" knowledge gaps emerge from incomplete information extraction. If such spurious gaps nevertheless remain, the researcher can identify them through its own gap analyses and instruct future digesters to address them through the research emphasis passed to those agents. Overall, these mechanisms should reduce the occurrence of false knowledge gaps.
    2. To further evaluate the effectiveness of knowledge-gap discovery, we are conducting an additional ablation study on a variant of WideResearcher with the knowledge-gap detection module removed. We will report the results as soon as the experiments are complete.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.