Event ·Official first-hand

Perplexity 与 NVIDIA 发布 JOLT：师生一体的后训练强化学习方法

Perplexity 和 NVIDIA 提出 JOLT，一种后训练阶段的强化学习方法，让同一个模型同时充当教师和学生进行联合训练。在 MATH 基准上，JOLT 达到了与 GRPO 相当的准确率，而处理训练 token 数量只有 GRPO 的约 1/6.5。该方法针对的是传统 RL 后训练中师生模型分离带来的训练开销问题。

Current read

Perplexity 和 NVIDIA 搞了个叫 JOLT 的训练方法，一个模型当老师又当学生，在 MATH 上准度追平 GRPO，训练 token 少了 6.5 倍。

11 sources47° AI heatLast updated 2026/10/22 04:56:31

Evidence

11 sources

Primary sourcemarktechpost

2026/10/22 04:56:31View source

Related source 1Yuchen Jin

2026/10/20 17:05:36View source

Related source 2arXiv cs.LG

2026/10/20 17:32:08View source

Related source 3berryxia

2026/10/20 11:34:11View source

Related source 4PyTorch

2026/10/20 16:53:50View source

Related source 5TestingCatalog

2026/10/20 22:02:03View source

Related source 6Teortaxes

2026/10/20 23:47:42View source

Related source 7SuperTechFans

2026/10/21 00:16:33View source

Related source 8rohanpaul_ai

2026/10/21 13:21:10View source

Related source 9CoreWeave

2026/10/21 17:32:54View source

Related source 10Merve Noyan

2026/10/22 06:54:10View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story