研究人员开发了一种名为AdaPrefix-GRPO的新技术，以改进语言模型在复杂推理任务上的训练。该方法在训练过程中自适应地调整提供给模型的参考解前缀量，旨在将成功率保持在梯度信号最强的约50%左右。一旦训练完成，模型无需此辅助即可解决问题，在具有挑战性的数学问题上显示出显著的准确性提升，尤其对较小的模型而言。 AI
影响 该方法可以显著提高小型语言模型在复杂推理任务上的性能，有可能减少对海量计算资源的需求。
排序理由 该集群描述了一种在arXiv论文中发布的新方法，用于改进AI模型在推理任务上的训练。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2607.07674v1 Announce Type: cross Abstract: Group Relative Policy Optimization (GRPO) stalls on a model's hardest problems: when no rollout in a group succeeds, the group-relative advantages vanish and the problem contributes no gradient, wasting the frontier examples we mo…
Group Relative Policy Optimization (GRPO) stalls on a model's hardest problems: when no rollout in a group succeeds, the group-relative advantages vanish and the problem contributes no gradient, wasting the frontier examples we most want to learn from. Prepending a correct prefix…
Group Relative Policy Optimization (GRPO) stalls on a model's hardest problems: when no rollout in a group succeeds, the group-relative advantages vanish and the problem contributes no gradient, wasting the frontier examples we most want to learn from. Prepending a correct prefix…