2d-strategy

Authors propose a 2D early-exit method that jointly trims layers and input sentences, yielding 1.4–2.3× extra speed-up on sentiment tasks across Llama 3.1/3.2, Gemma and Qwen models.