Cross-Domain Data Selection and Augmentation for Automatic Compliance Detection

The practical stakes here are higher than the abstract framing suggests: compliance detection systems trained on one regulatory body's language (say, GDPR) often degrade when applied to a different regime (say, financial services rules), and the wrong data selection approach can actively make performance worse, not just fail to improve it.

This connects most directly to the MIT Technology Review piece on 'Making AI operational in constrained public sector environments,' which flagged that government and regulated-industry deployments face governance constraints that generic model training ignores. The data selection problem studied here is essentially the technical underside of that governance challenge: you cannot simply fine-tune on whatever data is available when the target domain carries legal weight. The LLM judge reliability work covered around the same period ('Diagnosing LLM Judge Reliability') is also relevant, since compliance detection pipelines often use model-as-judge architectures where consistency failures compound transfer errors. Together, these papers sketch a picture of legal AI that is still working through foundational reliability problems before deployment questions even arise.

Watch whether any of the four selection strategies tested here get adopted or cited in compliance-focused benchmarks tied to specific regulatory corpora (like EU AI Act conformity assessments) within the next 12 months. Adoption there would signal the field is moving from lab results toward deployment-grade methodology.