斯坦福与Together AI提出混合推理Agent，性能提升10.4%
Banger paper from Stanford, Together AI, and colleagues.
Agent推理范式的重要突破，用混合机制显著降低成本并提升性能，值得研究Agent架构的同学关注。
Banger paper from Stanford, Together AI, and colleagues.
来自斯坦福大学、Together AI 及同事们的精彩论文。
Why is this an interesting paper?
为什么这篇论文有趣？
Today's LLM agents must write out every reasoning step as text before they can search, run code, or take action.
当前的 LLM 智能体在搜索、运行代码或采取行动之前，必须将每一个推理步骤以文本形式输出。
Those steps cost tokens and can take time.
这些步骤会消耗 token 并耗费时间。
This paper proposes Hybrid Reasoning Agents, which reason internally in latent space and switch to text only to call a tool or give an answer.
本文提出了混合推理智能体（Hybrid Reasoning Agents），它们在潜在空间（latent space）内部进行推理，仅在调用工具或给出答案时才切换为文本模式。
A small module called HybridLink moves the model between the two modes, and a working memory keeps both kinds of context across turns. Tool calls stay in plain text, and the latent steps can still be inspected.
一个名为 HybridLink 的小型模块负责在两种模式之间切换模型，工作记忆则在多轮对话中保持两类上下文。工具调用仍以纯文本形式进行，而潜在空间的步骤仍可被检查。
Training happens in two stages. SFT teaches the model to work with HybridLink, then RL teaches it when to think in latent space and when to act in text.
训练分为两个阶段。SFT（监督微调）教模型如何使用 HybridLink，随后 RL（强化学习）教模型何时在潜在空间中思考，何时以文本形式行动。
Across 11 benchmarks in math, science, medicine, search, and agentic reasoning, it outperforms frontier LLM agents and agentic frameworks, both text-based and latent, by 10.4% on average.
在数学、科学、医学、搜索和智能体推理领域的 11 个基准测试中，其平均性能比前沿 LLM 智能体和智能体框架（包括基于文本的和基于潜在空间的）高出 10.4%。
It also uses 35.5% to 54.7% fewer tokens, runs 1.8x to 2.6x faster, and costs 2.7x to 3.7x less.
此外，它使用的 token 数量减少了 35.5% 至 54.7%，运行速度提高了 1.8 倍至 2.6 倍，成本降低了 2.7 倍至 3.7 倍。
The 14B model beats the strongest agentic baseline by 17.4% on agentic reasoning tasks.
14B 参数模型在智能体推理任务上比最强的智能体基线模型高出 17.4%。
https://arxiv.org/abs/2610.18861
Chat with Paper: https://academy.dair.ai/papers/hybrid-reasoning-agents-2610.18861
与论文对话：https://academy.dair.ai/papers/hybrid-reasoning-agents-2610.18861
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