斯坦福论文Prefix Sliding：长推理无需全量CoT，推理提速3倍
New Stanford paper Prefix Sliding shows that long reasoning does not need the fu…
长上下文推理是热点难题，这篇论文用极简的Prefix Sliding策略实现了3倍加速且几乎无损，对Agent长程思考极具参考价值，建议关注其工程落地潜力。
New Stanford paper Prefix Sliding shows that long reasoning does not need the full chain of thought in memory: keep the task prefix and recent tokens, and inference can run about 3x faster without retraining.
斯坦福大学最新论文《Prefix Sliding》表明，长推理过程无需在内存中保留完整的思维链：只需保留任务前缀和最近的 token，即可在不重新训练的情况下将推理速度提升约 3 倍。
Long reasoning gets expensive because full attention makes every new token look back over an ever-growing chain of thought.
长推理之所以昂贵，是因为全注意力机制会让每个新 token 都回溯审视不断增长的思维链。
Prefix Sliding keeps the fixed prefix with task and tool instructions plus a sliding window of recent reasoning, dropping older middle tokens as it goes.
Prefix Sliding 保留了包含任务和工具指令的固定前缀，以及一个滑动窗口内的近期推理内容，并在过程中丢弃较旧的中间 token。
Their attention analysis points in the same direction: the prefix and latest tokens receive most attention, while intermediate reasoning gets little.
他们的注意力分析也指向同一方向：前缀和最新的 token 获得了大部分注意力，而中间的推理过程获得的注意力很少。
On Qwen3-1.7B, a 4,096-token window scored 33.9% on AIME25 versus 34.2% with full attention, and the paper reports about 3x faster inference without retraining.
在 Qwen3-1.7B 模型上，4,096 个 token 的窗口在 AIME25 测试中得分为 33.9%，而使用全注意力机制时为 34.2%；论文报告称，在不重新训练的情况下推理速度提升了约 3 倍。
Once the window fills, each new token has constant attention cost instead of becoming more expensive with every step, which also enabled reinforcement-learning rollouts beyond 100,000 tokens.
一旦窗口填满，每个新 token 的注意力成本保持恒定，而不是随着每一步变得更加昂贵，这也使得强化学习 rollout 能够扩展到超过 100,000 个 token。
Prefix Sliding beat pure sliding windows, repeated summarization, and last-k deletion on the tested speed/accuracy tradeoff.
在测试的速度/准确性权衡方面，Prefix Sliding 优于纯滑动窗口、重复摘要和最后 k 个删除方法。
– arxiv. org/abs/2608.26070
– arxiv.org/abs/2608.26070
Title: "Prefix Sliding for efficient test-time scaling"
标题：《用于高效测试时扩展的 Prefix Sliding》
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