Researchers have built a system of AI agents that watches large language models while they are being trained, not just after.
The system, called MASCRDM (Multi-Agent System for Compliance Risk Detection and Mitigation), comes from a team describing the approach in a paper posted to arXiv. The researchers first wrote compliance rules drawn from existing AI laws, then built a dedicated compliance-specific LLM, guided by legal experts, to apply those rules. They broke the model under training into components and used a compliance knowledge graph to flag which parts carry the most risk. Multiple agents then monitor those points throughout training, sending alerts and suggestions to developers as problems emerge, rather than waiting for a finished model to misbehave.
That timing is the real point. Most compliance tooling today checks a model's inputs and outputs after training wraps, a fix that catches symptoms without touching what is happening inside the model. MASCRDM instead treats bias and discrimination as something to watch for continuously while the model is still being shaped, a meaningfully different approach for an industry that mostly treats compliance as a release-day checklist.
The researchers report better scores on discrimination and bias benchmarks without a meaningful hit to language performance, though a paper's own benchmark numbers are always the easiest ones to improve.