Kimi k1.5
Scaling Reinforcement Learning with LLMs
...The project emphasizes a simplistic yet powerful framework where the context window scales up to 128k tokens, enabling reasoning that resembles planning, reflection, and correction over a much longer sequence of data than typical models. By using techniques like partial rollouts to improve training efficiency and applying sophisticated policy optimization methods, the developers demonstrate that strong ability can emerge without relying on complex solutions like Monte Carlo tree search or value functions. Kimi-k1.5 is trained jointly on text and vision data, giving it true multimodal reasoning capabilities where it can interpret and generate content across modalities in a unified way.