Adaptive Swarm

Setup设置

Every agent here is Gemini 4 Argon. The task is to make gigatoken, Marcel Rød's Rust tokenizer library, encode faster on one CPU core with the DeepSeek-V3 tokenizer, while every token id stays the same.

这里的智能体全部是 Gemini 4 Argon。任务是在单个 CPU 核上，让 Marcel Rød 的 Rust 分词库 gigatoken 用 DeepSeek-V3 分词器编码得更快，同时输出的每个 token id 都不变。

• Score. The original's best time over the candidate's, three alternating runs each on 500 MB of OpenWebText.
• Correctness. Identical token ids on held-out, edge-case and random inputs.
• Budget. Three hours per run, one idea per submission, one 32-core machine shared by the run's agents. A coordinator agent assigns the work and does not count toward the number of agents.

• 分数。在 500 MB 的 OpenWebText 上，原始版本和候选版本交替各跑三次，用原始版本的最好用时除以候选版本的最好用时。
• 正确性。在留出数据、边界用例和随机输入上，token id 都与原始版本完全一致。
• 预算。每次运行三小时，每次提交一个想法，同一次运行的智能体共用一台 32 核机器。另有一个协调者智能体负责分配工作，不计入智能体数。

Reasoning is cheap; execution is expensive. In research and science more broadly, the costly step is running an experiment, such as a training run or a lab assay, rather than thinking of it. This task is a cheap stand-in for that setting, yet we keep its accounting: the main cost is the number of executions, not model tokens, and every curve below runs over submissions.

推理便宜，执行昂贵。在更广泛的研究和科学工作中，昂贵的一步是把实验跑出来，例如一次训练或一次实验室测定，而不是想出这个实验。这里的任务是这种情形的一个便宜替代，但我们沿用同样的算法：主要成本是执行次数，而不是模型 token，下面所有曲线的横轴都是提交次数。

In a standard swarm@n, the coordinator hands work to up to n agents, and all of them edit one shared copy of the code. A recursive swarm@n puts managers between the coordinator and the workers, much as Recursive Language Models (Zhang et al., 2025) call themselves on parts of a problem and Prime Agent (Prime Intellect, 2026) has agents spawn sub-agents for subtasks.

标准蜂群@n 中，协调者把工作分给至多 n 个智能体，大家改同一份代码。递归蜂群@n 在协调者和工作者之间加了管理者，类似 Recursive Language Models（Zhang 等，2025）让模型对问题的一部分递归调用自己，以及 Prime Agent（Prime Intellect，2026）让智能体为子任务派生子智能体。

pass@k splits the agents into k groups, each on its own isolated copy of the code; pass@4 × recursive swarm@2 is four isolated recursive swarms of two agents. An adaptive swarm changes its groups during the run. We built five schedules for eight agents, A to E.

pass@k 把智能体分成 k 组，每组一份相互隔离的代码；pass@4 × 递归蜂群@2 是 4 个相互隔离、各有 2 个智能体的递归蜂群。自适应蜂群在运行中改变分组。我们为 8 个智能体设计了五种排程，A 到 E。

Experiments实验

Answered已有结论How do standard and recursive swarms change with the number of agents?标准蜂群和递归蜂群随智能体数怎样变化？

Both beat one agent at every size. Their final speedups are highest at four to eight agents and lower at 16 and 32.

每个规模都好于单个智能体。最终加速比在 4 到 8 个智能体时最高，16 和 32 个时更低。

Answered已有结论Do eight agents in isolated groups beat one swarm on one shared copy of the code?8 个智能体分成相互隔离的组，会好过共用一份代码的蜂群吗？

No. Every split of eight agents into isolated groups ends below a swarm of eight on one shared copy.

不会。8 个智能体的每一种隔离拆法，最终都低于共用一份代码的 8 智能体蜂群。

Testing检验中Does changing the groups during a run beat a fixed swarm?在运行中改变分组，能好过固定的蜂群吗？

Testing检验中How do CORAL, Fork-and-Flush and Stateless Language Agents do on this task?CORAL、Fork-and-Flush 和 Stateless Language Agents 在这个任务上表现如何？

Testing检验中How far does one agent get in 64 submissions?单个智能体做到 64 次提交能走多远？

Takeaways核心结论

For now, organizing agents in a more flexible way seems to work well.

用一种更灵活的方式组织 agent 似乎在目前是不错的。

Related work相关工作

CORAL (Qu et al., 2026) runs long-lived agents that explore asynchronously and share a persistent memory, with isolated workspaces and a separate evaluator; four co-evolving agents brought Anthropic's kernel task from 1363 to 1103 cycles. Like the swarms here, its agents attack one optimization target in parallel, and it studies how they share what they learn rather than how they are grouped.

CORAL（Qu 等，2026）让长时间运行的智能体异步探索、共享一份持久记忆，各自有隔离的工作区，评测器单独运行；四个共同演化的智能体把 Anthropic 的 kernel 任务从 1363 个周期降到 1103 个。和这里的蜂群一样，它的智能体并行攻同一个优化目标；它研究的是智能体之间如何共享所学，而不是如何分组。

Stateless Language Agents (Zhang et al., 2026) keep the research state outside the agents and start every advisor and worker call from a fresh context built for its role, which gives the best final results on budgets up to 1B tokens. It also finds that more inference alone does not bring more progress: what matters is how the experiments are organized.

Stateless Language Agents（Zhang 等，2026）把研究状态放在智能体之外，每次调用 advisor 和 worker 都从一个按角色新建的上下文开始，在最多 1B token 的预算下得到最好的最终结果。它同样发现，单靠更多推理不会带来更多进展，关键在于怎么组织实验。

Fork-and-Flush (Cai et al., 2026) shows that independent runs of one agent settle in different idea basins, and periodically forks the agent into branches that keep the workspace but clear the conversation, then keeps the best branch. It changes the shape of one search during the run, as the adaptive swarms here do with groups of agents.

Fork-and-Flush（Cai 等，2026）发现同一个智能体的独立运行会停在不同的 idea basin 里；它定期把智能体分叉成几个分支，保留工作区、清空对话，再留下最好的分支。它在运行中改变一次搜索的形状，这里的自适应蜂群则是在运行中改变智能体的分组。

Experiment record实验记录

Citation

@misc{chai2026adaptiveswarm,
  title        = {Adaptive Swarm},
  author       = {Chai, Wenhao},
  year         = {2026},
  howpublished = {Blog post},
  url          = {https://wenhaochai.com/blogs/adaptive-swarm.html}
}