#dynamic length reward

'VL-Cogito uses a staged curriculum RL and adaptive length rewards to significantly boost multimodal reasoning on math, science and chart benchmarks, outperforming several prior MLLMs.'