-
Real-world validation — Test on actual clinical documents (with appropriate ethics approval)
-
Automated fix application — Close the loop from proposed fixes to implementation
-
Cross-domain transfer — Do discovered weaknesses generalize to non-clinical German text?
-
Active learning — Use disagreement analysis to prioritize human annotation
-
ReAct: Yao, S., et al. (2022). "ReAct: Synergizing Reasoning and Acting in Language Models." ICLR 2023. arXiv:2210.03629
-
Reflexion: Shinn, N., et al. (2023). "Reflexion: Language Agents with Verbal Reinforcement Learning." NeurIPS 2023. arXiv:2303.11366
-
Multi-Agent Systems: Han, J., et al. (2024). "LLM Multi-Agent Systems: Challenges and Open Problems." arXiv:2402.03578
-
Curriculum Learning: Bengio, Y., et al. (2009). "Curriculum Learning." ICML 2009.
- anoner — Custom Presidio fork with German healthcare recognizers
- Microsoft Presidio — PII detection framework
- LangGraph — Workflow orchestration framework