Refining and Reusing Annotation Guidelines for LLM Annotation

In three linesLLMs struggle to follow specialized conventions of gold-standard benchmarks. Authors propose an iterative moderation framework that reuses and refines annotation guidelines as an alignment mechanism. Testing on three biomedical NER tasks (NCBI Disease, BC5CDR, BioRED) with GPT, Gemini, DeepSeek confirms efficacy of guideline integration and reasoning-optimized models.

Summary generated by Claude — human-verified