We introduce Nomad, a system for autonomous data exploration and insight discovery. Given a corpus of documents, databases, or other data sources, users rarely know the full set of questions, hypotheses, or connections that could be explored. As a result, query-driven question answering and prompt-driven deep-research systems remain limited by human framing and often fail to cover the broader insight space. Nomad addresses this problem with an exploration-first architecture. It constructs an explicit Exploration Map over the domain and systematically traverses it to balance breadth and depth. It generates and selects hypotheses and investigates them with an explorer agent that can use document search, web search, and database tools. Candidate insights are then checked by an independent verifier before entering a reporting pipeline that produces cited reports and higher-level meta-reports. We also present a comprehensive evaluation framework for autonomous discovery systems that measures trustworthiness, report quality, and diversity. Using corpora of selected UN and WHO reports and arXiv papers on LLM agents, we show that Nomad produces reports with strong numeric grounding, higher overall quality and actionability than baselines, and more diverse insights over several runs. Nomad is a step toward autonomous systems that not only answer user questions or conduct directed research, but also discover which questions, research directions, and insights are worth surfacing in the first place.
Figures & tables
Figure 1: Query-driven systems (left) explore only the narrow slice of a corpus that a user thinks to ask about, producing a single report shaped by the initial question. Nomad (right) constructs an Exploration Map over the domain and systematically traverses it, generating and verifying hypotheses across the full breadth of the corpus. This exploration-first approach yields diverse, evidence-backed reports covering directions the user may never have thought to pursue.
Nomad
Financing Healthier Electric Transport
Cyclones, Surveillance, and Hunger
Malaria’s Disability Data Gap
Shifting Stocks, Strained Seas
Stateless Women’s Twin Exposure
Data Readiness Boosts Funding
Table 1: A few example report titles generated by each system on the UN and WHO documents in the bottom-up setting. Nomad produces topically diverse titles, whereas o3-deep-research and GPT Researcher converge on similar themes and framings.
Figure 2: Overview of the Nomad pipeline. The exploration map and topic selection feed hypothesis generation, which enters the explorer–verifier loop. Verified insights are reported and evaluated, with periodic synthesis in meta-reports.
Figure 3: Interim state of the Exploration Map after concept layer construction. Documents are connected to the concepts extracted from them. Concepts are disambiguated across documents to point to a unified concept. The red (blue) highlighted nodes become part of a single LLM call during insight potential evaluation of concept c7 ( c3 ). This results in the generation of hypothesis h7 ( h3 ) as a child of the concept node.
Concept Name : COVID-19
Generated Hypothesis : The COVID-19 pandemic set off a reinforcing feedback loop in low- and middle-income countries wherein (a) economic shocks and falling labour-income shares increased poverty-related risk factors for diabetes and reduced access to routine care; (b) the resulting rise in severe infections among diabetics prompted higher, often inappropriate, antibiotic use; and (c) this surge in antibiotic consumption accelerated antimicrobial resistance, which in turn raises treatment costs and further strains household and health-system finances. Breaking this loop requires integrated policies that simultaneously address NCD management, antibiotic stewardship, and social protection.
Reasoning : All four documents speak of COVID-19 as a shock that spills over into very different domains: 1. the clinical vulnerability of people with diabetes and the disruption of their care, 2. the obligation of Member States to keep diabetes services running despite the pandemic, 3. the rise in antimicrobial resistance (AMR) partly fuelled by inappropriate antibiotic use during COVID-19, and 4. the pandemic-triggered fall in the global labour-income share that deepens inequality. These statements are not merely parallel references; they outline a potential causal chain that links a health-system stressor (COVID-19) to (i) higher diabetes morbidity/mortality, (ii) increased antibiotic use and AMR, and (iii) worsening socioeconomic conditions. The possibility that economic hardship and reduced labour income further exacerbate diabetes prevalence (through poverty-related risk factors) and drive inappropriate antibiotic use (through weakly regulated informal markets) suggests a feedback loop. Investigating this loop could uncover leverage points where integrated economic, stewardship, and NCD policies break the cycle – an area that is currently under-explored and highly actionable, especially for LMICs. Hence the seed set can credibly launch a multi-sectoral research thread with game-changing policy implications.
Supporting Documents
D1. WHO Health Topics Diabetes A74 R4, p. 2. …increased across both periods; [2] Concerned that people living with noncommunicable diseases, including diabetes, have a higher risk of becoming severely ill or dying from coronavirus disease (COVID-19), and are among those most impacted by the COVID-19 pandemic; [3] Concerned also that complete or partial disruptions to diabetes prevention and control due to the COVID-19 pandemic, including in respect of early detection and diabetic complication management services, represent significant threats to the life and health of people living with diabetes; 1 See document A74/10 Rev.1. …
D2. WHO Health Topics Diabetes A74 R4, p. 4. …for developing type 2 diabetes; (5) to ensure a continued focus on maintaining a high level of treatment and care for all people, regardless of the COVID-19 pandemic, including for people living with diabetes, especially in low- and middle-income countries, recognizing that necessary diabetes prevention and control efforts are hampered by, inter alia, lack of universal access to quality, safe, effective, affordable essential health services, medicines, diagnostics and health technologies, as well as by a global shortage of qualified health workers; (6) to ensure that national …
D3. WHO Health Topics Antimicrobial Resistance A77-5, p. 3. …focus on drug-resistant bacterial infections. However, other infections, such as viral and fungal infections, may elicit the inappropriate use of antibiotics and are also considered. The pandemic of coronavirus disease (COVID-19), for example, contributed to antimicrobial resistance. As further evidence emerges on …
Table 2: Concept-layer hypothesis example: COVID-19. The hypothesis and reasoning are generated from the documents associated with the concept, and the decisive snippets are highlighted in blue .
Concept Name : ILO
Generated Hypothesis : Strategic public investment in expanding the health and care workforce—as advocated by the WHO-ILO-OECD ‘Working for Health’ action plan—can offset automation-induced declines in the labour-income share, reduce youth NEET rates (especially for women), and mitigate rising inequality over the next decade. Empirical modelling that integrates sectoral labour-intensity profiles, automation susceptibility indices, and macro labour-income-share trends can quantify the potential distributional and growth effects of such health-employment expansion.
Reasoning : The documents build a coherent narrative: (a) the global labour-income share has been falling; (b) automation-oriented technological change is a statistically significant contributor; (c) this erosion risks worsening inequality and youth NEET rates, especially for young women. A separate WHO resolution jointly adopted with ILO and OECD promotes large-scale investment in the health workforce - particularly community health workers - as a lever for ‘health employment and inclusive economic growth’. The ILO report itself notes that some forms of technological innovation can raise the labour-income share when they ‘create labour-intensive tasks’. Health and care services are well-known to be comparatively resistant to automation and are highly labour-intensive, offering significant employment opportunities for youth and women.
Hence, there is a non-trivial, cross-document connection: while automation is squeezing labour’s share in many sectors, coordinated global expansion of the health-care workforce (championed by the joint WHO-ILO-OECD plan) could offset that squeeze by generating labour-intensive, less-automatable jobs, simultaneously addressing NEET rates and gender gaps. This link is neither explicit in the ILO update nor in the WHO resolution, yet combining them suggests an actionable research avenue: can strategic investment in health employment act as a macro-level counterbalance to automation-induced inequality and youth unemployment?
Because the documents provide complementary empirical findings, theoretical framing (Acemoglu and Restrepo), and a concrete policy programme (Working for Health plan), they form a strong seed for high-impact, multidisciplinary research that could model and quantify such a compensatory strategy, explore financing mechanisms, and guide policy design.
Supporting Documents
D2. ILO World Employment Update Sep 2024, p. 3. …Outlook: September 2024 Update 3 While the decrease appears modest in percentage points (a reduction of 1.6 p.p. in two decades, of which 0.6 p.p. in the last 5 years), the effect is sizeable. This decline represents $2.4 trillion (in constant PPP) in labour income globally in 2024. Notwithstanding the distributional significance of …
Table 3: Concept-layer hypothesis example: ILO
Concept Name : 2018
Generated Hypothesis : Insufficient progress on the 2018 UNGA diabetes commitments—particularly inadequate insulin accessibility in low- and middle-income countries—has indirectly contributed to the continued global rise of antimicrobial resistance observed in 2018-2023 by increasing infection incidence and antibiotic consumption among people with poorly controlled diabetes. Integrated strategies that simultaneously expand diabetes care (insulin supply, lifestyle intervention) and AMR stewardship could yield greater impact than siloed approaches.
Reasoning : The three documents are thematically different (diabetes, antimicrobial resistance policy, and AMR surveillance) yet all anchor key data or policy commitments to the year 2018. [Document 1] cites the 2018 UNGA resolution 73/2 that called for intensified primary-care diabetes action. [Document 2] cites the 2018 UNEP resolution on AMR and urges stronger One-Health AMR plans. [Document 3] presents hard surveillance data showing that, despite policy efforts, median annual AMR for several critical pathogens has risen steadily from the 2018 baseline through 2023. Because diabetes substantially increases infection risk and antibiotic consumption (e.g., diabetic foot infections, urinary-tract infections), inadequate diabetes control—exacerbated by poor insulin access highlighted in [Document 1]—can indirectly fuel inappropriate or prolonged antibiotic use, thereby contributing to AMR. Meanwhile, [Document 3] reports that AMR is highest where surveillance (and, by implication, health-system capacity) is weakest—often the same low- and middle-income countries that [Document 1] notes have poor insulin access. Thus the documents can be linked to form a non-trivial research thread: failures to meet the 2018 diabetes-care commitments may be an unrecognised driver undermining the 2015-2022 AMR action plans, helping to explain why AMR trends since 2018 are worsening despite policy pledges. Exploring this intersection could reveal synergistic interventions (e.g., increasing affordable insulin and glycaemic control to curb infection incidence, thereby reducing antibiotic pressure and slowing AMR).
Supporting Documents
D1. Diabetes A74 R4 - p. 3. …segments of the population; [1] Reaffirming also our commitment in United Nations General Assembly resolution 73/2 (2018) to further strengthen efforts to address diabetes as part of universal health coverage through intensified interventions at the primary health care level, including in low- and middle-income countries, on prevention and control of diabetes; Emphasizing the importance of prevention …
D2. Antimicrobial Resistance A72 R5, p. 2. …and promote their prudent use; Recalling FAO resolution 4/2015 on antimicrobial resistance, World Organisation for Animal Health (OIE) resolution No. 36 (2016) on combating antimicrobial resistance through a One Health approach: actions and OIE strategy, and the UNEP resolution UNEP/EA.3/Res.4 (2018) on environment and health; Noting the importance of providing opportunities for Member States to engage meaningfully with and provide input into reports, recommendations, and relevant actions from WHO, FAO, and OIE, together with UNEP, and from the Interagency Coordination Group on Antimicrobial Resistance aimed at combating antimicrobial resistance; Reaffirming the global commitment to combat antimicrobial resistance with continued, high-level political efforts as a coordinated international community, emphasizing the critical need to accelerate Member States’ development and implementation of their national action plans with a One Health approach , 1. WELCOMES the new tripartite …
D3. Global Health GLASS AMR Surveillance Report 2025, p. 11. …report higher levels of AMR. The frequency of AMR is highest in countries with low surveillance coverage. In fact, there is a strong inverse correlation between a country’s AMR surveillance coverage and its reported median AMR (Pearson correlation coefficient, r=−0.74,P<0.0001 , Fig. 4). This pattern may reflect both …
Table 4: Concept-layer hypothesis example: 2018
Figure 4: The full exploration map: A Topic Tree is constructed on top of the concept layer, and concepts are connected to their closest leaf topic node. The Topic Tree is traversed breadth-first for exploration.
Figure 5: A real exploration map from the WHO Research Analyst instance. The Topic Tree (circle nodes) is constructed over the concept layer (rectangles), with each concept connected to its closest leaf topic node via dashed lines. Note the non-uniform branching factor and depth across different branches of the tree.
Figure 6: Running example of exploration map construction for the WHO Research Analyst use case. The figure shows both top-down construction from the research goal alone (left) and bottom-up construction from a provided corpus (right). While the top-down map is more general and covers more diverse topics (using the LLMs world knowledge about topics relevant to the goal), the bottom-up map is more detailed and specific to the topics found in the small set of documents provided - demonstrating the relevance and relative advantages of both approaches. Moreover, the bottom-up map has non-uniform branching factor and depth, which is a direct consequence of the insight potential evaluation step during concept layer construction, which awards individual concepts with suitable insight potential with hypotheses before the Topic Tree construction, thus potentially resulting in topic nodes with more (or fewer) such concept nodes in their subtree.
Figure 7: Seed-guided topic insertion during Web search expansion of the exploration map. At the first level, T2 is selected based on relevance to the seed set. At the second level, since no existing child matches the seed, a new topic T6 is proposed and inserted as a child of T2 .
Tree Path : Methods & Data > FEW Nexus Modeling > Integrated Assessment
Relevant Retrieved Seeds (Web Search results based on Tree Path)
S10
Title : Towards an Inclusive Water-Energy-Food Nexus Framework - UNU . Snippet : The framework enhances access to WEF resources by all, gender equality and social … URL : unu.edu/inweh/event/science-talk-towards-inclusive-water-energy-food-nexus-framework
Jul 13, 2024
S14
Title : 4 global risks to look out for in the post-pandemic era . Snippet : The post-pandemic era is being shaped by heightened global risk and unpredictable shock events. … URL : weforum.org/stories/2024/08/4-global-risks-to-look-out-for-in-the-post-pandemic-era
Aug 19, 2024
S15
Title : After Shock to After Strategy—a Resilient System or New … . Snippet : The COVID-19 pandemic produced a foundational jolt to health care. … URL : jamanetwork.com/journals/jamanetworkopen/fullarticle/2841286
Nov 12, 2025
S23
Title : Integrating UN data and AI for Accelerated Recovery - UN-GGIM . Snippet : Integrating UN data and AI for accelerated recovery. GIS provides solutions for hazards, crops and planning. … URL : ggim.un.org/meetings/GGIM-committee/12th-Session/side_events/UNGN_UNDP.pdf
S5
Title : Nexus approach to enhance water-energy-food security …- Nature . Snippet : Model-based nexus assessments can compare climate, socio-economic and demographic scenarios. … URL : nature.com/articles/s44168-025-00308-4
Dec 16, 2025
Table 5: Hypothesis-generation example for a sampled leaf topic in exploration. The upper half lists the subset of relevant Web searches which contributed to the generated hypotheses. The lower half shows the resulting hypotheses and their respective scores on all 3 dimensions. The selected hypothesis is boldfaced.
Figure 8: Explorer–verifier loop. The Explorer loops through multiple rounds of observe-reason-act workflow making multiple tool calls at each step. Once an insight is generated, the Verifier evaluates it by splitting it into sub-claims and validating each. If the validation passes, the system moves on to report generation. Otherwise, the feedback is used by the explorer to refine the insight. Sources referenced at each step of explorer/verifier are collected into a citation database, and utilized by the report generation.
Figure 9: AI-generated summary of a real explorer–verifier trace from the WHO Research Analyst example (initial hypothesis from Table 3 ).
Section #
Section Type
Section Plan
Report Title: Care Economy’s Triple Dividend
1
Text
The Automation–Inequality Nexus and the Care-Economy Solution Summarises ILO WESO evidence on automation pressure and youth NEET stagnation, recaps the WHO 11 million health-worker shortfall, and frames the care economy as a lever for inclusive growth. Introduces the twin challenges of falling labour-income shares and persistent youth NEET rates, then positions health- and care-workforce expansion as a cross-sectoral response that addresses both trends.
2
Table
Consolidated Evidence for a Triple Dividend Plans one compact evidence table with the 11 million shortfall, the 3.4 employment multiplier, the 299 million care-jobs estimate, the female youth-NEET gap, and low automation exposure for care work. Aggregates headline numbers from WHO, World Bank, and ILO to ground the report’s argument in concrete, cross-agency evidence.
3
Chart
Automation Exposure by Sector: Where Health and Care Stand Plans a sector comparison chart that places human health and social work near the low-risk end of automation exposure using task-based ILO and OECD evidence. Visualises why shifting employment toward health and care can hedge against AI-driven job displacement, reinforcing the third element of the triple dividend.
4
Measurement
Quantified Impact Scenario: Meeting the 11 Million Gap by 2030 Turns the evidence into scenario metrics: direct jobs, indirect jobs from the multiplier, a projected female youth-NEET reduction, and a labour-income-share buffer. Translates evidence into forward-looking metrics to show the scale of possible macro benefits, bridging the gap between descriptive data and policy relevance.
5
Recommendation
Stakeholder-Specific Policy Recommendations Plans stakeholder-specific recommendations for governments, multilateral development banks, education and training systems, and private employers. Provides actionable guidance that operationalises the report’s insight for the main decision-makers.
Table 6: Generated outline, excluding executive summary, for the WHO Research Analyst example in Table 3 . Blue bands show section titles, amber bands show section descriptions, and green bands show section-level throughlines. The report generator fills each section with content one at a time, yet remains tied to the overall narrative using this explicit outline as a scaffold.
Figure 10: Overview of the report generation workflow. The finalized insight and citation database are first used to draft a report structure. Section filling and citation auditing then iteratively refine this draft into a structured, evidence-grounded report. The audited report is finally transformed into a compact poster that highlights the main insight and key supporting evidence.
Figure 11: An overview of the meta-report generation pipeline. Inputs from the insight pool ( I1,…,IN ), goal from this Nomad’s instance (green), feed into the processing stages below. Purple boxes denote pipeline stages: relevance distillation produces K highlights ( H1,…,HK ), which undergo entity-based clustering into thematic groups ( C1,…,Cj ). Each cluster is then refined through thematic analysis, analysis enrichment, and analysis condensation before final report generation.
Figure 12: Example generated meta-report for the running WHO analyst example.
Tag t
Score w(t)
ref
1.0
sec_ref
0.8
report_ref
0.7
prev_section
0.5
explorer
0.5
misattributed_section
0.4
Table 7: Numeric grounding evaluation tags and their corresponding scores.
Verdict v
Score Q(v)
SUPPORTED
1.0
PARTIAL
0.5
UNCITED
0.3
UNVERIFIABLE
0.25
UNSUPPORTED
0.0
Table 8: Factuality evaluation verdicts and their corresponding scores.
Keyphrase
Class
Score
Demonstrates a clear and logical causal chain.
Very Good
+2
Defines limitations and boundary conditions explicitly.
Very Good
+2
Provides strong and comprehensive citation coverage.
Very Good
+2
Shows mostly sound reasoning throughout.
Good
+1
Briefly acknowledges limitations.
Good
+1
Supports most claims with evidence.
Good
+1
Table 9: Quality-evaluation attributes and scores for analytical quality and originality.
Keyphrase
Class
Score
Provides broad, multi-angle coverage.
Very Good
+2
Thoroughly explores the central hypothesis.
Very Good
+2
Aligns strongly with the topic and stated goal.
Very Good
+2
Covers most of the relevant dimensions.
Good
+1
Delivers strong topic coverage.
Good
+1
Presents generally relevant and focused content.
Good
+1
Table 10: Quality-evaluation attributes and scores for coverage and actionability.
Keyphrase
Class
Score
Maintains an engaging and professional tone.
Very Good
+2
Features a clear and well-structured structure.
Very Good
+2
Uses highly readable and professional language.
Very Good
+2
Adopts a generally engaging tone.
Good
+1
Demonstrates mostly well-structured organization.
Good
+1
Communicates clearly and effectively.
Good
+1
Table 11: Quality-evaluation attributes and scores for presentation.
Score
Meaning
4
Almost no repetition. The sections are effectively independent.
3
Slight repetition. There are one or two minor overlaps, but the sections remain clearly distinct.
2
Some repetition. There are multiple overlapping points, and the sections are only moderately distinct.
1
Severe repetition. A large amount of content is repeated, and distinctness is low.
0
Excessive repetition. The two sections are nearly entirely repetitive.
Table 12: Distinctness evaluation scores and their interpretation.
WHO
arXiv
Dimension
GPT Researcher
o3-deep-research
Nomad
GPT Researcher
o3-deep-research
Nomad
# Reports
35
35
35
35
35
35
Trustworthiness
Numeric Grounding (%)
34.1
65.7
60.2
45.3
64.7
65.9
Factuality
27.6
74.5
50.9
36.1
82.4
45.5
Quality
Table 13: Evaluation results across all dimensions for the bottom-up setting in the WHO and arXiv domains. Each configuration is capped at 35 reports (verified reports for Nomad). For quality, we report the overall aggregated score and the five dimensions.
WHO
arXiv
Dimension
GPT Researcher
o3-deep-research
Nomad
GPT Researcher
o3-deep-research
Nomad
# Reports
35
35
35
35
35
35
Trustworthiness
Numeric Grounding (%)
42.4
67.6
65.8
33.5
78.0
67.5
Factuality
47.8
72.9
59.5
40.0
79.4
68.2
Quality
Table 14: Evaluation results across all dimensions for the top-down setting in the WHO and arXiv domains. Each configuration is capped at 35 reports (verified reports for Nomad). For quality, we report the overall aggregated score and the five dimensions.
Figure 13: Multi-dimensional evaluation profiles in the WHO and arXiv domains under bottom-up and top-down settings. Each axis is on a 0 – 100 scale, with diversity multiplied by 100 for comparability. Nomad has the highest overall quality, actionability, and diversity in all four combinations, while the trustworthiness tradeoffs vary by baseline and setting.
Figure 14: Average total numeric claims (light bars) and grounded numeric claims (dark bars) per report under the 35-report cap. Percentages show the share of numeric claims that are grounded. These raw counts are distinct from the weighted Numeric Grounding scores in Tables 13 and 14 , which weight each numeric claim according to where its supporting evidence is found.
Domain
Setting
Explorer
NG
Fact.
Overall
Analyt.
Cov.
Act.
Pres.
Orig.
Dist.
Div.
WHO
Bottom-up
azure-o3
60.2
50.9
60.1
56.2
67.6
46.3
83.9
46.7
61.1
0.5315
WHO
Bottom-up
gpt5
66.7
65.6
66.0
61.0
77.5
56.0
85.5
49.9
57.6
0.4944
arXiv
Bottom-up
azure-o3
65.9
45.5
63.1
61.1
69.4
52.6
85.0
47.4
63.6
0.3972
arXiv
Bottom-up
gpt5
68.4
53.7
72.0
68.7
79.9
66.3
86.1
59.0
57.0
0.3541
WHO
Top-down
azure-o3
65.8
59.5
61.6
61.1
68.8
45.6
84.8
47.5
56.7
0.5376
WHO
Top-down
gpt5
67.6
60.7
63.0
58.9
76.1
47.2
85.4
47.1
52.1
0.4727
Table 15: Explorer-model ablation. NG denotes numeric grounding; Fact., Overall, Analyt., Cov., Act., Pres., Orig., Dist., and Div. denote factuality, overall quality, analytical quality, coverage, actionability, presentation, originality, distinctness, and diversity. The verifier is fixed to Llama-3.3-70B-Instruct .
Domain
Setting
Verifier
NG
Fact.
Overall
Analyt.
Cov.
Act.
Pres.
Orig.
Dist.
Div.
WHO
Bottom-up
Llama-3.3-70B
60.2
50.9
60.1
56.2
67.6
46.3
83.9
46.7
61.1
0.5315
WHO
Bottom-up
o4-mini
59.4
54.1
62.1
58.1
71.0
47.3
84.9
49.1
59.6
0.5217
WHO
Bottom-up
gemma4-31b
60.5
48.5
61.3
58.7
70.4
46.8
85.7
44.8
58.8
0.5235
arXiv
Bottom-up
Llama-3.3-70B
65.9
45.5
63.1
61.1
69.4
52.6
85.0
47.4
63.6
0.3972
arXiv
Bottom-up
o4-mini
67.0
46.7
65.3
63.7
71.7
51.8
84.8
54.4
56.5
0.4026
arXiv
Bottom-up
gemma4-31b
62.9
40.0
62.1
60.7
68.8
48.3
85.0
47.7
64.7
0.4091
Table 16: Verifier-model robustness ablation with the explorer fixed to azure-o3 . Metric abbreviations follow Table 15 .
Domain
Setting
Configuration
NG
Fact.
Overall
Analyt.
Cov.
Act.
Pres.
Orig.
Dist.
Div.
WHO
Bottom-up
Full Nomad
60.2
50.9
60.1
56.2
67.6
46.3
83.9
46.7
61.1
0.5315
WHO
Bottom-up
No hypothesis generation
64.8
57.4
59.9
60.7
72.5
44.6
82.9
38.5
53.1
0.3847
arXiv
Bottom-up
Full Nomad
65.9
45.5
63.1
61.1
69.4
52.6
85.0
47.4
63.6
0.3972
arXiv
Bottom-up
No hypothesis generation
65.4
44.4
62.4
65.9
71.8
48.8
84.7
40.6
61.2
0.2925
WHO
Top-down
Full Nomad
65.8
59.5
61.6
61.1
68.8
45.6
84.8
47.5
56.7
0.5376
WHO
Top-down
No hypothesis generation
68.9
55.4
62.9
62.5
81.9
49.7
85.6
34.8
54.8
0.3509
Table 17: Hypothesis-generation ablation with the explorer and verifier fixed to the anchor configuration. Metric abbreviations follow Table 15 .
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Nomad
Embedded Guardrails Elevate LLMs
Hierarchical Memory Slashes Hallucinations
Beyond Benchmarks: Probing Agent Reliability
Capability Tokens: Securing Agentic AI
Adversarial Critics Slash Hallucinations
Causal Explainability Boosts Agent Safety
Appendix
Table 18: A few example report titles generated by each system on the arXiv papers about LLM-based agents in the bottom-up setting. Nomad produces topically diverse titles, whereas o3-deep-research and GPT Researcher converge on similar themes and framings.
Autonomous research agents use language models and tools to search literature, design and run experiments, analyze results, and support scientific discovery.
Scientific discovery in data-rich domains is currently constrained by human bandwidth: the growth in the volume and complexity of real-world data far outpaces the rate at which researchers can read, reason, and synthesize. Recent LLM-based multi-agent systems have begun to automate portions of the research cycle, but they target hypothesis generation in settings where validation cannot itself be automated, and each run is one-shot, with no mechanism for findings to accumulate or steer subsequent inquiry. This paper introduces AutoKD, a multi-agent framework for autonomous knowledge discovery that is both computational and cumulative, allowing validated findings to persist and inform subsequent inquiry. Six coordinated LLM agents collaborate in an open-ended discovery loop, where accepted findings are stored in a persistent insight graph that serves as both long-term memory and an exploration-steering mechanism. We evaluate AutoKD on three diverse datasets from two perspectives: Open-ended Quality against published findings, and Conditioned Quality via literature-derived queries. Across both evaluation perspectives, AutoKD covers known findings and surfaces substantive discoveries that complement human-driven research. Our code is available at https://github.com/GeQinwen/AutoKD.
Qinwen Ge, Bo Ni, Haowei Fu +3
Vanderbilt University · Air Force Research Laboratory
We present Odin, the first production-deployed graph intelligence engine for autonomous discovery of meaningful patterns in knowledge graphs without prior specification. Unlike retrieval-based systems that answer predefined queries, Odin guides exploration through the COMPASS (Composite Oriented Multi-signal Path Assessment) score, a novel metric that combines (1) structural importance via Personalized PageRank, (2) semantic plausibility through Neural Probabilistic Logic Learning (NPLL) used as a discriminative filter rather than generative model, (3) temporal relevance with configurable decay, and (4) community-aware guidance through GNN-identified bridge entities and inter-community affinity scores. This multi-signal integration, particularly the bridge scoring mechanism, addresses the "echo chamber" problem where graph exploration becomes trapped in dense local communities. We formalize the autonomous discovery problem, prove theoretical properties of our scoring function, and demonstrate that beam search with multi-signal guidance achieves O(b⋅h) complexity while maintaining high recall compared to exhaustive exploration. To our knowledge, Odin represents the first autonomous discovery system deployed in regulated production environments (healthcare and insurance), demonstrating significant improvements in pattern discovery quality and analyst efficiency. Our approach maintains complete provenance traceability -- a critical requirement for regulated industries where hallucination is unacceptable.
Deep research with Large Language Model (LLM) agents is emerging as a powerful paradigm for multi-step information discovery, synthesis, and analysis. However, existing approaches primarily focus on unstructured web data, while the challenges of conducting deep research over large-scale structured databases remain relatively underexplored. Unlike web-based research, effective data-centric research requires more than retrieval and summarization and demands iterative hypothesis generation, quantitative reasoning over structured schemas, and convergence toward a coherent analytical narrative. In this paper, we present DataSTORM, an LLM-based agentic system capable of autonomously conducting research across both large-scale structured databases and internet sources. Grounded in principles from Exploratory Data Analysis and Data Storytelling, DataSTORM reframes deep research over structured data as a thesis-driven analytical process: discovering candidate theses from data, validating them through iterative cross-source investigation, and developing them into coherent analytical narratives. We evaluate DataSTORM on InsightBench, where it achieves a new state-of-the-art result with a 19.4% relative improvement in insight-level recall and 7.2% in summary-level score. We further introduce a new dataset built on ACLED, a real-world complex database, and demonstrate that DataSTORM outperforms proprietary systems such as ChatGPT Deep Research across both automated metrics and human evaluations.
Shicheng Liu, Yucheng Jiang, Sajid Farook +3
Computer Science Department, Stanford University, Stanford, CA