A carregar...
Publicação
Designing LLM Agents for Output Checking in Research Data Centers
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 675.61 KB | Adobe PDF |
Orientador(es)
Resumo(s)
Research data centers increasingly face a tension between rapidly growing demand for microdata-driven research and strict confidentiality obligations. Output checking—reviewing tables, models, and descriptive statistics before release from safe centers-remains a key safeguard but is labor-intensive and difficult to scale. This paper presents a governed, semi-automated output-checking architecture that integrates three layers: (i) deterministic rule-of-thumb disclosure checks, (ii) a machine-learning classifier trained on historical release decisions to support triage, and (iii) a large language model (LLM) agent that performs principles-based synthesis and produces structured explanations for auditors and researchers. Using an archival corpus of output requests and released/blocked outputs from a national statistical office, we build a metadata and standardization pipeline for heterogeneous datasets and outputs, engineer disclosure-relevant features, and evaluate a prototype that routes low-risk cases quickly while escalating ambiguous cases for human control. We evaluate the architecture retrospectively on historical output-checking records and distinguish three validation targets: deterministic detection of obvious disclosure risks, ML prediction of institutional outcomes, and LLM-based evidence synthesis for audit support. The contribution is a governance-oriented architecture and empirical prototype for routing and explanation, not an autonomous release mechanism: all unsupported, ambiguous, or high-risk cases remain subject to professional output-checker review.
Descrição
Ashofteh, A., Carvalho, R., & Campos, P. (2026). Designing LLM Agents for Output Checking in Research Data Centers: Disclosure Risk and Output Validation. In 2026 IEEE 50th Annual Computers, Software, and Applications Conference (COMPSAC) (pp. 189-195). (Proceedings of the Annual Computer Software and Applications Conference). IEEE Computer Society. https://doi.org/10.1109/COMPSAC69091.2026.00035
Palavras-chave
Large language models AI agents secure data safe rooms microdata access statistical disclosure control output validation
Contexto Educativo
Citação
Editora
IEEE Computer Society