OpenAI最新模型Astra用上字节架构：Transformer同一层反复算

2026-09-02 12:51

Share To

据媒体报道，OpenAI 即将发布的 Astra 模型采用了循环推理架构 recurrent depth（也称 looped transformer）。

与普通 Transformer 每生成一个 Token 固定穿过各层网络不同，Astra 允许同一段信息反复经过同一组层，多轮计算后再输出。字节跳动去年已公开类似路线。

其 Seed 团队发布的 Ouro 是一种 Looped Language Model，同样让一组 Transformer 层循环运行，将更多计算内置于模型内部，使较小模型通过增加推理计算量接近更大模型的效果。

该架构带来安全隐忧：部分推理发生在内部隐藏状态，人类无法看到完整文字记录，更难通过思维链检查模型是否违规。

OpenAI 因此限制了 Astra 对 recurrent depth 的使用程度，保留可读的思维链，并计划加入额外的 CoT 监控。

Disclaimer: Contains third-party opinions, does not constitute financial advice

Recommended Reading

Amazon announces a multi-billion dollar investment in Missouri to build a data center campus, expected to create over 400 long-term positions

06-16

Elon Musk's current personal net worth has risen to the range of $1.1 trillion to $1.2 trillion.

06-15

OpenAI and Anthropic Employees Cash Out $14 Billion in Five Years

06-15

Tencent Involved in Investment in Lin Junyang's AI Lab, Former Head of Alibaba's Qwen, Valued at $2 Billion

06-15

Binance will delist the CVC/USDC, RPL/USDC, RVN/USDC, and XAI/USDC leveraged trading pairs on June 19

06-15

U.S. Security Experts Unite to Urge Repeal of Anthropic Ban, Warn That Restricting Defensive Tools Will Lose the AI Race Against China

06-15

Musk vows SpaceX revenue to reach $1 trillion by 2030

06-15

Amazon announces a multi-billion dollar investment in Missouri to build a data center campus, expected to create over 400 long-term positions

Elon Musk's current personal net worth has risen to the range of $1.1 trillion to $1.2 trillion.

OpenAI and Anthropic Employees Cash Out $14 Billion in Five Years

Tencent Involved in Investment in Lin Junyang's AI Lab, Former Head of Alibaba's Qwen, Valued at $2 Billion

Binance will delist the CVC/USDC, RPL/USDC, RVN/USDC, and XAI/USDC leveraged trading pairs on June 19

U.S. Security Experts Unite to Urge Repeal of Anthropic Ban, Warn That Restricting Defensive Tools Will Lose the AI Race Against China

Musk vows SpaceX revenue to reach $1 trillion by 2030