SpecMind

SpecMind is a spectrum intelligence dataset released alongside the paper:

“SpecMind: Enabling Spectrum Intelligence via Multi-Agent Hybrid Retrieval-Augmented Generation.”

This repository contains two complementary components:

1. a heterogeneous spectrum knowledge dataset, and
2. SpecBench, an expert-designed benchmark for evaluating retrieval-augmented generation (RAG) systems in the spectrum domain.

The dataset supports research on retrieval and reasoning across structured, textual, and graph-oriented spectrum information.

Dataset Components

Spectrum Knowledge Dataset

The knowledge dataset contains three types of spectrum-domain information:

• License data — structured spectrum licensing records with numerical and categorical attributes.
• Proceeding documents — FCC proceeding comments and reply comments involving multiple stakeholders and regulatory topics.
• Regulatory documents — codified spectrum rules and policies, including FCC regulatory text.

These sources are used by SpecMind with modality-aware retrieval mechanisms: SQL-based retrieval for license data, graph-based retrieval for proceedings, and dense retrieval for regulatory text.

The proceeding collection used in the accompanying work includes:

• FCC 19-38
• FCC 24-72
• FCC 25-59
• FCC 22-352
• FCC 23-158
• FCC 23-232
• NTIA National Spectrum Strategy

SpecBench

SpecBench is a Q&A benchmark for evaluating RAG systems over heterogeneous spectrum data.

It contains 450 curated question-answer pairs based on realistic spectrum-analysis tasks and evidence from the accompanying knowledge sources.

SpecBench evaluates three core capabilities:

• Noise Robustness — retrieving the correct evidence despite irrelevant or noisy information.
• Information Integration — combining evidence across multiple documents, records, or data sources.
• Negative Rejection — recognizing when sufficient evidence is unavailable and avoiding unsupported answers.

Question Categories

SpecBench contains five major categories:

Category	Data Source	Main Capability
Proceeding	Proceeding documents	Noise robustness / information integration
License	License records	Noise robustness / information integration
Regulation	FCC Title 47	Information integration
Compound	Multiple data sources	Information integration
Unanswerable	No valid evidence	Negative rejection

Proceeding and license questions include both single-evidence and multi-evidence tasks.

Compound questions are further divided into:

• Parallel: evidence can be retrieved independently from different sources and combined in the final answer.
• Sequential: results from one retrieval step are required to formulate or constrain subsequent retrieval steps.

The benchmark distribution is:

Category	Percentage
Proceeding	31.1%
License	31.1%
Regulation	13.3%
Compound	14.4%
Unanswerable	10.0%

Regulation questions also include samples from the WiLL benchmark.

Repository Structure

SpecMind/
├── knowledge/
│   ├── proceedings/
│   ├── licenses/
│   └── regulations/
│
└── benchmark/
    └── SpecBench

The knowledge/ directory contains the spectrum-domain resources used for retrieval.

The benchmark/ directory contains SpecBench questions, reference answers, and associated evaluation information.

Intended Uses

This dataset is intended for research on:

• retrieval-augmented generation,
• multi-agent and agentic RAG,
• spectrum policy and regulatory question answering,
• heterogeneous information retrieval,
• structured and graph-based retrieval,
• cross-source reasoning, and
• hallucination and negative-rejection evaluation.

SpecMind Framework

The accompanying SpecMind framework uses specialized agents for different knowledge sources:

• License Agent — SQL-based retrieval over structured license data.
• Proceeding Agent — graph-based retrieval over proceeding documents.
• Regulation Agent — dense retrieval and reranking over regulatory text.

A Supervisor Agent coordinates these components for multi-source spectrum reasoning.

Code is available at:

https://github.com/swdong01/SpecMind

Licensing and Source Documents

This repository contains both original dataset components created by the authors and materials derived from or redistributed from external spectrum-domain sources.

Unless otherwise specified, the original components created by the authors, including SpecBench question-answer annotations, benchmark organization, metadata, and evaluation labels, are released under the Creative Commons Attribution 4.0 International License (CC BY-NC 4.0).

External source materials, including FCC/NTIA proceeding documents, regulatory text, spectrum licensing records, and any derived records or excerpts based on those sources, remain subject to the terms, conditions, and public-domain status of their original sources. Redistribution through this repository does not replace, modify, or supersede the original source terms.

Users are responsible for consulting the original sources for applicable licensing, usage conditions, attribution requirements, and legal or regulatory constraints.

Citation

@misc{dong2026specmind,
  title         = {SpecMind: Enabling Spectrum Intelligence via Multi-Agent Hybrid Retrieval-Augmented Generation},
  author        = {Dong, Songwei and Lu, Bingyan and Kienlen, Makayla and Laneman, J. Nicholas and Shen, Cong},
  year          = {2026},
  eprint        = {2609.00427},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2609.00427}
}

Acknowledgment

This work was supported in part by SpectrumX, the National Science Foundation (NSF) Spectrum Innovation Center, operated under Cooperative Agreement by the University of Notre Dame. The project was conceived during a SpectrumX center meeting and subsequently received a seed fund to carry out the research; we thank NSF and SpectrumX for the support.

The authors would like to thank Zhiyu Shen, Yuxi Chen, Omkar Mujumdar, Yankai Peng, Hasan Nazim Bicer, Christopher Wahl, Solbee Kang, and Lucas Scholler in the Notre Dame Wireless Institute for their valuable contributions to data collection and the evaluation process. In addition, we thank Caleb Reinking, Le Li Kruczek, Connor Howington, and Paul Brenner in the Notre Dame Center for Research Computing for their contributions to code optimization and the implementation of the system user interface.