MechLens: Late Crystallization of Factual Knowledge Explains Intervention Effectiveness in Language Models
Researchers have identified a phenomenon called "Late Crystallization" in large language models, where factual knowledge primarily emerges in the final layers rather than gradually across all layers. This finding, observed across multiple model families like Pythia, Gemma, and Llama-3.1, suggests that factual recall is concentrated towards the end of the model's processing. The study also proposes a new intervention principle based on this crystallization and introduces a spectrum distinguishing between computable and memorized knowledge. AI
IMPACT Reveals that LLMs store factual knowledge late, potentially guiding future model design and intervention strategies for accuracy.