Where transformers commit to an answer
Researchers found a distinct layer in transformer models where answer predictions lock in — and it's consistent across models and fine-tuning.
Researchers found a distinct layer in transformer models where answer predictions lock in — and it's consistent across models and fine-tuning.