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.
Figures & tables
Figure 1: Overview of the DataSTORM system. A complete research workflow consists of three stages: (1) a warm-start module powered by internet research, (2) a multi-agent exploration module, and (3) final report generation.
Figure 2: Overview of the Final Report Generation Module
Method
Insight-level
Summary-level
GPT-4o judge
Qwen3-30B judge
GPT-4o judge
Qwen3-30B judge
Previous SOTA
AgentPoirot (gpt-5)
47.1
49.9
46.6
51.5
Our Approach
DataSTORM (gpt-5)
61.9
69.3
52.5
58.7
Table 1: Insight- and summary-level recall (%) on InsightBench ( Sahu et al., 2025 ) , evaluated under two LLM judges. DataSTORM outperforms the state-of-the-art AgentPoirot consistently under both judges. Qwen3-30B refers to Qwen/Qwen3-30B-A3B-Instruct-2507 .
Ref-Induced
RACE Eval Framework ( Du et al., 2025 )
DB Use
Approach
Avg Match
Overall
Comp.
Depth
Inst.
Read.
Ratio
OpenAI DR (MCP)
48.5%
46.1
45.6
42.8
49.3
49.3
30.4%
OpenAI DR (CSV)
51.2%
46.8
46.9
43.9
49.4
49.5
23.3%
DataSTORM
61.8%
52.6
52.5
52.2
53.5
52.5
66.4 %
Table 2: Comparison of OpenAI Deep Research and DataSTORM on the ACLED benchmark using reference-induced criteria matching and the RACE evaluation framework from Du et al. (2025) . All scores are evaluated using GPT-5.
Rubric
OpenAI DR (CSV)
DataSTORM
Win% (lose%)
Public Impact
3.58
3.60
25.0% (20.0%)
Public Interest
3.30
3.50
32.5% (20.0%)
Institutional Significance
3.45
3.58
30.0% (17.5%)
Conflict and Stakes
3.43
3.83
45.0% (22.5%)
Originality
2.21
3.05 †
61.5% (10.3%)
Human Consequences
2.75
2.88
37.5% (35.0%)
Table 3: Human ratings of newsworthiness dimensions for reports generated by DataSTORM and OpenAI DR (CSV) ( N=40 reviewer-topic evaluations; N=39 for Originality because one evaluator marked one topic as N/A ). Ratings use a 1–5 scale. Evaluators could assign N/A when an article did not provide sufficient basis for judgment; such cases were excluded from the corresponding statistics. We report mean ratings and the pairwise win rate of DataSTORM. † indicates a significant difference ( p<0.05 ) between DataSTORM and OpenAI DR (CSV) in a paired t-test.
Figure 3: Pairwise preference rates in human evaluation. For each comparison, the bar shows the percentage of evaluations preferring the left-side system versus the right-side system.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Consent page shown to participants before the human evaluation. Identifying information has been redacted for anonymous review.
Figure 5: Screenshot of the custom web interface used for human evaluation. Participants reviewed the reference article and system outputs, assigned rubric-based newsworthiness ratings, and provided pairwise comparisons and optional free-form comments.
Table 4: Ablation results for DataSTORM on the 10 ACLED topics. All scores judged by GPT-5.
Axis
Correct
Unsure
Incorrect
Faithfulness to cited result table
86%
4%
10%
Retrieval correctness of SQL query
94%
6%
0%
Appendix
Table 5: Manual verification of 50 numerical, SQL-backed claims sampled from 18 DataSTORM reports on the ACLED benchmark.
Dimension
What it checks
Example inconsistency
Actor / entity matching
Whether the same actor or group is matched the same way
One query searches only actor1 and actor2 for “Fano,” while another also searches assoc_actor_* and notes
Geography
Whether the same place is normalized the same way
One query uses admin1 = ’Amhara’ , while another uses country = ’Ethiopia’ AND admin1 = ’Amhara’
Temporal
Whether the same time period uses the same date window
One query uses event_date >= ’2023-04-01’ , while another uses BETWEEN ’2023-04-01’ AND ’2024-12-31’
Event taxonomy
Whether the same event concept is defined consistently
“Political violence” is defined using disorder_type in one query but event_type in another
Counting / aggregation
Whether the same quantity is counted the same way
One query reports COUNT(*) (number of events) while another reports SUM(fatalities) (death toll) for the same magnitude claim
Population scope
Whether the same base population is filtered the same way
One query includes country = ’Sudan’ , while another omits it for the same regional claim
Appendix
Table 6: The six ACLED-relevant dimensions of the query consistency metric, and the kind of inconsistency each one catches.
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · Also with University of Chinese Academy of Sciences. · National University of Singapore, Singapore, Singapore