On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance
The research investigated how LLM performance in annotation tasks relates to model-internalized knowledge versus user instructions. It found that performance is best predicted by Definition-Specific Familiarity (DSF)—how well a model's internal concept matches the task definition—rather than simple text memorization. This suggests focusing on conceptual alignment during setup is more important for reliable results.