研究人员引入了 Kernelized Advantage Estimation (KAE) 来通过强化学习增强大型语言模型 (LLM) 的推理能力。KAE 解决了现有方法（如 Proximal Policy Optimization 和 GRPO）的局限性，这些方法要么计算开销高，要么需要过多的采样。通过利用经典的非参数统计方法，特别是核平滑，KAE 旨在以更少的每次提示推理轨迹来实现准确的值和梯度估计。这种方法在资源受限的环境中尤其有益，有望改善 LLM 的策略优化。 AI
影响 提供了一种在资源受限的情况下，通过强化学习改进 LLM 推理能力的更具计算效率的方法。
排序理由 这是一篇介绍 LLM 推理新方法的学术论文。
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
arXiv:2604.28005v1 Announce Type: cross Abstract: Recent advances in large language models (LLMs) have increasingly relied on reinforcement learning (RL) to improve their reasoning capabilities. Three approaches have been widely adopted: (i) Proximal policy optimization and advan…
Recent advances in large language models (LLMs) have increasingly relied on reinforcement learning (RL) to improve their reasoning capabilities. Three approaches have been widely adopted: (i) Proximal policy optimization and advantage actor-critic rely on a deep neural network to…