Researchers have developed a new method for training large language models offline, without requiring online interactions. The approach, called Future Policy Approximation (FPA), addresses a major challenge in offline learning by reweighting gradients to avoid suppressing critical tokens. FPA estimates the future policy through logit-space extrapolation and outperforms strong baselines on various reasoning and code-generation benchmarks. --- Why it matters: This matters because it enables more efficient training of large language models, which can be computationally expensive and unstable in online settings. By improving offline learning methods, researchers can develop more accurate and robust models for complex tasks like mathematical reasoning and code generation. Source: https://arxiv.org/abs/2509.19893