Description
本期播客深入探讨了DAPO（解耦裁剪与动态采样策略优化）算法，这是一个在Qwen2.5-32B基础模型上实现AIME 2024测试50分的先进大规模强化学习系统。我们详细讨论了其四项关键技术：Clip-Higher、动态采样、词元级策略梯度损失和超长奖励修正，以及它们如何解决熵塌陷、梯度消失、长CoT场景下的学习不平衡和奖励噪声等问题，并介绍了其开放源代码、训练代码和精心处理的数据集对社区的贡献。
Audio
Featured in this Episode
No persons identified in this episode.
Transcription
This episode hasn't been transcribed yet
Help us prioritize this episode for transcription by upvoting it.
0
upvotes
Popular episodes get transcribed faster
Other recent transcribed episodes
Transcribed and ready to explore now
Eric Larsen on the emergence and potential of AI in healthcare
10 Dec 2025
McKinsey on Healthcare
Reducing Burnout and Boosting Revenue in ASCs
10 Dec 2025
Becker’s Healthcare -- Spine and Orthopedic Podcast
Dr. Erich G. Anderer, Chief of the Division of Neurosurgery and Surgical Director of Perioperative Services at NYU Langone Hospital–Brooklyn
09 Dec 2025
Becker’s Healthcare -- Spine and Orthopedic Podcast
Dr. Nolan Wessell, Assistant Professor and Well-being Co-Director, Department of Orthopedic Surgery, Division of Spine Surgery, University of Colorado School of Medicine
08 Dec 2025
Becker’s Healthcare -- Spine and Orthopedic Podcast
NPR News: 12-08-2025 2AM EST
08 Dec 2025
NPR News Now
NPR News: 12-08-2025 1AM EST
08 Dec 2025
NPR News Now