Meta Superintelligence Labs 发布
Great paper from Meta Superintelligence Labs on compounding inference for long-h…
多智能体协作推理的新范式，GitSwarm 利用版本控制实现知识累积与复用，数据扎实且效果显著，值得 Agent 研究者关注。
Great paper from Meta Superintelligence Labs on compounding inference for long-horizon agents.
Meta 超级智能实验室发表了一篇关于长程智能体复合推理的优秀论文。
One clear approach that's working with large multi-agent systems is persistent memory.
在大型多智能体系统中，一种行之有效的方法是持久化记忆。
Meta researchers describe a concrete approach for this in this paper.
Meta 研究人员在这篇论文中描述了实现这一目标的具体方法。
You'll find that inference-time compute usually goes to separate attempts, so most runs ignore partial work and failed ideas.
你会发现，推理时的计算资源通常被分配给独立的尝试，因此大多数运行会忽略部分工作和失败的想法。
Here is how they solve this problem:
以下是他们解决这个问题的方法：
GitSwarm gives hundreds of asynchronous agents one shared Git repository.
GitSwarm 为数百个异步智能体提供了一个共享的 Git 仓库。
Each contribution is a commit that records which earlier commits, on any branch, it builds on. There is no central planner. Each worker decides whether to explore, verify, refine, or combine existing work.
每次贡献都是一个提交记录，它记录了该提交基于哪些分支上的早期提交进行构建。没有中央规划器。每个工作者自行决定是探索、验证、完善还是组合现有工作。
With GPT-5.5 Codex workers, it solves all 30 IMOProofBench-Advanced problems in one run. On ProgramBench, it scores 79.4%, against 65.1% for a single Codex agent prompted to keep going at a similar budget.
使用 GPT-5.5 Codex 工作者，它在一次运行中解决了 IMOProofBench-Advanced 的全部 30 个问题。在 ProgramBench 上，其得分为 79.4%，而单个 Codex 智能体在类似预算下被提示继续运行时得分仅为 65.1%。
Score keeps rising with more workers, from 73.1% at 100 to 79.4% at 500. Later workers build on 94.7% of published contributions, and 70.1% of failed research commits are cited by later accepted work.
随着工作者数量增加，分数持续上升，从 100 名工作者的 73.1% 提升至 500 名工作者的 79.4%。后期工作者利用了 94.7% 的已发布贡献，且 70.1% 的失败研究提交被后续通过的工作所引用。
Paper: https://arxiv.org/abs/2610.04862
论文：https://arxiv.org/abs/2610.04862
Chat with Paper: https://academy.dair.ai/papers/gitswarm-decentralized-compounding-inference-2610.04862
与论文对话：https://academy.dair.ai/papers/gitswarm-decentralized-compounding-inference-2610.04862
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