Google DeepMind提出Recirculation：无需重训提升Tr
You don't often see one-word titles in AI papers.
这篇论文展示了一种无需重训即可增强Transformer推理能力的创新架构思路，性能提升显著且计算开销可控，对Agent长程推理有启发意义。
You don't often see one-word titles in AI papers.
在 AI 论文中，你很少见到单字标题。
That aside, strong recommend this paper from Google DeepMind.
撇开这一点不谈，强烈推荐 Google DeepMind 的这篇论文。
I think this is an interesting training-free approach to evolve model architectures by leveraging the model itself to inform architectural modifications.
我认为这是一种有趣的无需训练的架构演化方法，它利用模型自身来指导架构修改。
Something like this could also inspire even more robust recursive self-improvement approaches.
类似这样的思路还可能启发更具鲁棒性的递归自我改进方法。
Approach details below:
方法细节如下：
A feedforward transformer can only update its internal state as many times as it has layers. Long generations need more updates than that, so chain-of-thought ends up doing basic state tracking in text.
前馈 Transformer 只能根据其层数更新内部状态相同的次数。生成长文本需要更多的更新，因此思维链（Chain-of-Thought）最终是在文本中进行基本的状态跟踪。
Recirculation adds recurrence at inference time.
循环机制（Recirculation）在推理时引入了递归性。
The model feeds activations back through itself during prefill, which lets it act like a dynamical system and track belief states without any retraining.
模型在预填充（prefill）阶段将激活值反馈回自身，这使其能够像动力系统一样运作，并在无需重新训练的情况下跟踪信念状态。
Generation cost stays flat. All the serial work happens in prefill.
生成成本保持平稳。所有串行工作都在预填充阶段完成。
On the Gemma3 family, the adaptive variant cuts perplexity 23% and lifts GSM8k accuracy 21%, with the original weights frozen and only light hyperparameter tuning.
在 Gemma3 系列上，自适应变体将困惑度降低了 23%，并将 GSM8k 准确率提升了 21%，且原始权重被冻结，仅进行了轻量级的超参数调整。
Paper: https://arxiv.org/abs/2608.17981
论文：https://arxiv.org/abs/2608.17981
Track more trending AI papers in our academy: https://academy.dair.ai/
在我们的学院中追踪更多热门 AI 论文：https://academy.dair.ai/
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力