On March 31, 2026, a six-person research team submitted Nomad: Autonomous Exploration and Discovery to arXiv, presenting a system that autonomously generates and investigates hypotheses from large document or database corpora — without requiring users to define what questions to ask. The authors — Bokang Jia, Samta Kamboj, Satheesh Katipomu, Seung Hun Han, Neha Sengupta, and Andrew Jackson — argue that query-driven and prompt-driven research systems are fundamentally constrained by the limits of human framing, and that Nomad is designed to address that structural gap.
- Nomad constructs an explicit Exploration Map to systematically traverse a knowledge domain, balancing breadth against depth in hypothesis generation and investigation.
- An explorer agent invokes document search, web search, and database query tools; a separate independent verifier checks every candidate insight before it enters the reporting pipeline.
- Evaluated on a corpus of selected UN and WHO reports, Nomad produced more trustworthy, higher-quality, and more diverse insights than baseline systems across multiple independent runs.
- The paper also introduces a new evaluation framework that measures autonomous discovery systems across three dimensions: trustworthiness, report quality, and diversity.
What Happened
The Nomad paper introduces what the authors call an “exploration-first architecture” for AI-assisted research. Rather than responding to user-defined queries, the system first constructs an Exploration Map of the knowledge domain and then systematically traverses it to identify which hypotheses are worth investigating. The team demonstrated the system on a corpus of selected United Nations and World Health Organization reports, producing cited reports and higher-level meta-reports without any user-specified research questions guiding the process.
Why It Matters
Current AI research tools — including retrieval-augmented generation systems and prompt-driven deep research products — depend on users knowing what to ask. The authors describe this as systems that “remain limited by human framing and often fail to cover the broader insight space.” When a dataset is large or unfamiliar, users cannot enumerate the full set of questions, hypotheses, or connections that might be relevant — meaning entire categories of insight go undetected.
Nomad’s contribution is upstream of retrieval: it automates the question-generation stage itself, deciding which directions are worth investigating before any retrieval or synthesis begins. This is architecturally distinct from deep-research products that accept a user prompt and then find supporting evidence.
Technical Details
At the core of Nomad is the Exploration Map, an explicit structured representation of the domain that guides the system’s traversal strategy. The map is designed to balance breadth — covering a wide range of potential insight areas — against depth, pursuing specific hypotheses in detail where the domain warrants it. An explorer agent carries out investigations by invoking document search, web search, and database query tools depending on the available data sources.
A key architectural decision is the use of a separate, independent verifier module that evaluates candidate insights before they enter the reporting pipeline. This separation of generator and verifier roles is intended to improve trustworthiness by preventing unverified claims from reaching the final output. The system ultimately produces both granular cited reports and higher-level meta-reports that synthesize findings across multiple investigations.
The evaluation framework the team developed measures three dimensions: trustworthiness, report quality, and diversity. On the UN and WHO corpus, Nomad outperformed baseline systems across all three metrics. Critically, the diversity gains held across several independent runs, indicating the system does not consistently surface the same set of insights — a failure mode that would undermine the value of repeated autonomous exploration.
Who’s Affected
The system is most directly relevant to knowledge workers, policy analysts, and researchers who operate with large, unstructured document corpora — including scientific literature, institutional reports, or internal enterprise documentation. The UN and WHO corpus used in evaluation represents the kind of sprawling, multi-domain dataset where manual auditing for overlooked connections is impractical at scale.
Developers building autonomous data analysis or research pipelines may find the Exploration Map and verifier architecture worth examining, given the paper’s treatment of trustworthiness as a measurable, optimizable property rather than a qualitative claim.
What’s Next
The authors describe Nomad as “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.” As of publication, no external release or production deployment has been announced. The evaluation corpus — UN and WHO reports — represents a narrowly scoped test case, and the paper does not demonstrate performance across diverse domain types such as scientific literature, financial data, or legal documents.
The evaluation framework itself is a new contribution without independent replication. Confirming whether the trustworthiness and diversity gains generalize beyond the tested corpus will require follow-up work from the authors or external researchers applying the framework to different data sources.
