#LLM Inference

When the Answer Arrives Late: Early-Exit Decoding in Modern LLMs