We propose Bingo, an RL framework that enhances efficient reasoning in LLMs through advanced length-based reward design. It incorporates two key mechanisms: a significance-aware length reward that gradually guides the model to reduce only insignificant tokens, and a dynamic length reward that…