1/5 As part of our work on improving the efficiency of our LLM online-RL training pipelines, we cut policy update step time by ~70% by introducing a model-agnostic padding minimization method.

→ View original post on X — @ai21labs,