diff --git a/.github/workflows/pages.yml b/.github/workflows/pages.yml index a91bb30..78dd539 100644 --- a/.github/workflows/pages.yml +++ b/.github/workflows/pages.yml @@ -35,6 +35,14 @@ jobs: pip install "mkdocs-material>=9.5" "mkdocs-static-i18n>=1.2" mkdocs build --strict -d _site/docs + # ── Live demo: arbor replay --demo --html -> _site/demo.html ───── + # Regenerated from source on every deploy so the hosted, zero-install + # demo always matches the shipped code (no committed artifact to go stale). + - name: Build live demo page + run: | + pip install . + arbor replay --demo --html --no-open --out _site/demo.html + # ── Project page: Vite/React -> _site (root) ───────────────────── - uses: actions/setup-node@v4 with: diff --git a/README.md b/README.md index 4e11b96..cbba20d 100644 --- a/README.md +++ b/README.md @@ -3,13 +3,15 @@

-# Toward Generalist Autonomous Research via Hypothesis-Tree Refinement +

🌳 Arbor

+

The autonomous research agent that beats Claude Code and Codex by 2.5× on the same compute budget

Paper GitHub Project Page + Live Demo Docs Discussions License: Apache 2.0 @@ -19,18 +21,29 @@ English | 简体中文

-**Arbor is an autonomous research agent that turns a long-horizon objective into a -cumulative search.** Give it a benchmark and a goal; it proposes hypotheses, edits -code, runs real experiments, learns from the results, and keeps the improvements that -hold up on held-out data. Instead of one-shot attempts that forget what failed, Arbor -grows a **hypothesis tree**: every idea becomes a branch — pruned if it fails, -harvested if it works — and insights propagate back so later ideas start smarter. +

+ Give Arbor a benchmark and a goal. It proposes hypotheses, edits code, runs real + experiments, and keeps only the gains that survive held-out data — growing a + hypothesis tree instead of forgetting what failed. +

+ +> **▶️ Try it in 30 seconds — no API key, no config:** +> +> ```bash +> pip install arbor-agent && arbor replay --demo # watch the hypothesis tree grow live +> ``` +> +> Or **watch it right now in your browser** — nothing to install: **[▶️ Live Demo](https://RUC-NLPIR.github.io/Arbor/demo.html)**. + +### 🏆 One controller, six tasks — wins the held-out test on all of them + +| Task | Metric | Claude Code | Codex | **Arbor** | +| --- | --- | :---: | :---: | :---: | +| BrowseComp | acc ↑ | 53.33 | 50.00 | **67.67** | +| Terminal-Bench 2.0 | pass ↑ | 71.70 | 73.59 | **77.36** | +| Math-Reasoning Data | gap ↑ | 8.33 | 6.25 | **20.83** | -For more details, visit our [project page](https://RUC-NLPIR.github.io/Arbor/) -and read the [paper](https://arxiv.org/pdf/2606.11926). For a more detailed usage manual, -see our [documentation](https://RUC-NLPIR.github.io/Arbor/docs/). 🧭 You can also -choose the [CLI or Skill version](#-cli-and-skill-versions) depending on your -environment and workflow. +Plus **86.36% Any-Medal on MLE-Bench Lite** (GPT-5.5). → [See all six tasks](#-results) · [project page](https://RUC-NLPIR.github.io/Arbor/) · [paper](https://arxiv.org/pdf/2606.11926) · [docs](https://RUC-NLPIR.github.io/Arbor/docs/) ## 🎬 Demo @@ -39,10 +52,10 @@ https://github.com/user-attachments/assets/49c1a306-d2e9-49d6-9c83-65e38a62df30 ## 📣 News -- **2026-06** — **Built-in literature search & idea novelty checks.** Arbor can now ground its research in prior work via the public [alphaXiv](https://www.alphaxiv.org) API — zero config, no search endpoint or key. Novelty-check any idea before you build it with `arbor idea-check ""`, or let the Coordinator vet every new branch automatically. See [Literature Search & Novelty Checks](#-literature-search--novelty-checks). 🔎 -- **2026-06** — Arbor was featured by [VentureBeat](https://venturebeat.com/), one of the leading tech media outlets in the US: ["New AI optimization framework beats Claude Code and Codex by 2.5x on the same compute budget"](https://venturebeat.com/orchestration/new-ai-optimization-framework-beats-claude-code-and-codex-by-2-5x-on-the-same-compute-budget). 📰 -- **2026-06** — Arbor's native CLI runtime and Agent Skill Suite (Codex / Claude Code) are released. 🚀 -- **2026-06** — The Arbor paper is released on [arXiv](https://arxiv.org/abs/2606.11926). 🎉 +- **2026-06-22** — **Built-in literature search & idea novelty checks.** Arbor can now ground its research in prior work via the public [alphaXiv](https://www.alphaxiv.org) API — zero config, no search endpoint or key. Novelty-check any idea before you build it with `arbor idea-check ""`, or let the Coordinator vet every new branch automatically. See [Literature Search & Novelty Checks](#-literature-search--novelty-checks). 🔎 +- **2026-06-18** — Arbor was featured by [VentureBeat](https://venturebeat.com/), one of the leading tech media outlets in the US: ["New AI optimization framework beats Claude Code and Codex by 2.5x on the same compute budget"](https://venturebeat.com/orchestration/new-ai-optimization-framework-beats-claude-code-and-codex-by-2-5x-on-the-same-compute-budget). 📰 +- **2026-06-12** — Arbor's native CLI runtime and Agent Skill Suite (Codex / Claude Code) are released. 🚀 +- **2026-06-11** — The Arbor paper is released on [arXiv](https://arxiv.org/abs/2606.11926). 🎉 ## 💡 Why Arbor diff --git a/README.zh-CN.md b/README.zh-CN.md index 3087b01..c9e15f4 100644 --- a/README.zh-CN.md +++ b/README.zh-CN.md @@ -3,13 +3,15 @@

-# 基于假设树的面向通用自主科研方法(Toward Generalist Autonomous Research via Hypothesis-Tree Refinement) +

🌳 Arbor

+

在相同算力预算下,效果超越 Claude Code 与 Codex 2.5× 的自主科研智能体

Paper GitHub Project Page + Live Demo Docs Discussions License: Apache 2.0 @@ -19,20 +21,35 @@ English | 简体中文

-**Arbor 是一个自主科研智能体,可以把长周期目标转化为持续累积的搜索过程。** 给它一个基准 -(benchmark)和一个目标,它会提出假设、修改代码、运行真实实验、从结果中学习,并保留那些在 -留出(held-out)数据上经得起验证的改进。不同于“一次性尝试、过后即弃”的做法,Arbor 会逐步生长出一棵 -**假设树**:每个想法都是一根分支,失败则剪枝,成功则保留;洞见会沿树反向传播,让后续想法 -从更可靠的起点出发。 +

+ 给 Arbor 一个基准和一个目标。它会提出假设、修改代码、运行真实实验,只保留经得起留出 + 数据验证的改进——生长出一棵假设树,而不是过后即弃、忘记失败的教训。 +

+ +> **▶️ 30 秒上手——无需 API key,无需配置:** +> +> ```bash +> pip install arbor-agent && arbor replay --demo # 实时观看假设树的生长过程 +> ``` +> +> 或者**现在就在浏览器里看**——无需安装任何东西:**[▶️ 在线 Demo](https://RUC-NLPIR.github.io/Arbor/demo.html)**。 + +### 🏆 一个控制器,六项任务——全部赢下留出测试集 + +| 任务 | 指标 | Claude Code | Codex | **Arbor** | +| --- | --- | :---: | :---: | :---: | +| BrowseComp | acc ↑ | 53.33 | 50.00 | **67.67** | +| Terminal-Bench 2.0 | pass ↑ | 71.70 | 73.59 | **77.36** | +| Math-Reasoning Data | gap ↑ | 8.33 | 6.25 | **20.83** | -更多详情,请访问我们的[项目主页](https://ruc-nlpir.github.io/Arbor/)并阅读[论文](https://arxiv.org/pdf/2606.11926)。如需详细的使用说明,请参阅[文档](https://ruc-nlpir.github.io/Arbor/docs/)。🧭 你也可以根据自己的环境和工作流选择使用 [CLI 版本或技能套件版本](https://claude.ai/chat/e7121091-ce2c-4970-a60f-16b54c453729#-cli-与技能套件版本)。 +外加 **MLE-Bench Lite 上 86.36% Any-Medal**(GPT-5.5)。→ [查看全部六项任务](#-实验结果) · [项目主页](https://ruc-nlpir.github.io/Arbor/) · [论文](https://arxiv.org/pdf/2606.11926) · [文档](https://ruc-nlpir.github.io/Arbor/docs/) ## 📣 最新动态 -- **2026-06** — **内置文献检索与想法新颖性审查。** Arbor 现在可以通过 [alphaXiv](https://www.alphaxiv.org) 公共 API 把研究建立在已有工作之上——零配置,无需搜索端点或密钥。动手前用 `arbor idea-check "<你的想法>"` 审查任意想法的新颖性,或让 Coordinator 自动为每个新分支把关。详见[文献检索与新颖性审查](#-文献检索与新颖性审查)。🔎 -- **2026-06** — Arbor 被美国知名科技媒体 [VentureBeat](https://venturebeat.com/) 报道:[《New AI optimization framework beats Claude Code and Codex by 2.5x on the same compute budget》](https://venturebeat.com/orchestration/new-ai-optimization-framework-beats-claude-code-and-codex-by-2-5x-on-the-same-compute-budget)。📰 -- **2026-06** — Arbor 原生 CLI 运行时与智能体技能套件(Codex / Claude Code)正式发布。🚀 -- **2026-06** — Arbor 论文在 [arXiv](https://arxiv.org/abs/2606.11926) 发布。🎉 +- **2026-06-22** — **内置文献检索与想法新颖性审查。** Arbor 现在可以通过 [alphaXiv](https://www.alphaxiv.org) 公共 API 把研究建立在已有工作之上——零配置,无需搜索端点或密钥。动手前用 `arbor idea-check "<你的想法>"` 审查任意想法的新颖性,或让 Coordinator 自动为每个新分支把关。详见[文献检索与新颖性审查](#-文献检索与新颖性审查)。🔎 +- **2026-06-18** — Arbor 被美国知名科技媒体 [VentureBeat](https://venturebeat.com/) 报道:[《New AI optimization framework beats Claude Code and Codex by 2.5x on the same compute budget》](https://venturebeat.com/orchestration/new-ai-optimization-framework-beats-claude-code-and-codex-by-2-5x-on-the-same-compute-budget)。📰 +- **2026-06-12** — Arbor 原生 CLI 运行时与智能体技能套件(Codex / Claude Code)正式发布。🚀 +- **2026-06-11** — Arbor 论文在 [arXiv](https://arxiv.org/abs/2606.11926) 发布。🎉 ## 💡 为什么选择 Arbor diff --git a/docs/roadmap.md b/docs/roadmap.md index 23a57b2..1aa95f1 100644 --- a/docs/roadmap.md +++ b/docs/roadmap.md @@ -164,6 +164,52 @@ sources behind each idea) and a per-run cost breakdown. --- +## Direction 4 — Adoption, DX & community + +The first three directions grow what Arbor *does*. This one lowers the barrier +between a curious visitor and a first successful run, and makes the project's +momentum visible. The goal is a shorter path to the "aha" moment, not more +surface area. + +### 4.1 Live, zero-install demo + +`arbor replay --demo --html` already emits a self-contained, dependency-free page +of a real run. Publish that page (e.g. on GitHub Pages alongside the docs) and +link it from the top of the README, so a visitor can watch the hypothesis tree +grow **without installing anything**. One recorded run, refreshed when the +dashboard changes. + +### 4.2 Examples gallery + +Today `examples/algotune_knn` is the only end-to-end task a newcomer can run. +Grow this into a small gallery so different audiences can self-identify, each +with a copy-pasteable command and a short asciinema recording: + +- a Kaggle / MLE-style task (pairs with the `mle_kaggle` plugin), +- a prompt / harness-engineering task, and +- a small training / fine-tuning task. + +Keep the bar at "runs in minutes on a laptop or a free key", reusing the +zero-config discipline of `algotune_knn`. + +### 4.3 Regular releases and a changelog + +Versions are already derived from `v*` tags via setuptools-scm, so the cost of a +release is low. Cut a release on a predictable cadence with human-readable notes, +and keep a `CHANGELOG.md` (or GitHub Releases as the source of truth) so the +project's progress is legible from the outside. + +### 4.4 Public roadmap and issue board + +Surface this document as a public board (GitHub Projects) and tag tracked work +with `good first issue` / `help wanted`. Several threads are already +contribution-shaped — growing the [benchmark zoo](#21-benchmark-zoo-organized-by-domain-format-tooling-shipped-collection-growing) +to 3–5 packs ([2.1](#21-benchmark-zoo-organized-by-domain-format-tooling-shipped-collection-growing)) +and adding domain plugins ([2.2](#22-plugin-gallery)) — so opening them up turns +readers into contributors. + +--- + Have an idea or want to own one of these threads? Open a [discussion](https://github.com/RUC-NLPIR/Arbor/discussions) or see [Contributing](contributing.md). diff --git a/docs/roadmap.zh.md b/docs/roadmap.zh.md index be3cfdc..5060c77 100644 --- a/docs/roadmap.zh.md +++ b/docs/roadmap.zh.md @@ -70,7 +70,7 @@ ## 方向二 —— 外部资源 -### 2.1 按 domain 划分的 benchmark zoo 🚧 *(格式与工具已完成;集合扩充中)* +### 2.1 按 domain 划分的 benchmark zoo 🚧 *(格式与工具已完成;集合扩充中)* {#sec-2-1} 一个经过筛选、统一格式的任务集合,按领域分组(如 CV、NLP、时序、优化),每个任务用一篇已 发表论文的结果作为要超越的 baseline。它以 `arbor-zoo/` 放在仓库里,每个基准一个文件夹, @@ -106,7 +106,7 @@ 复用 `plugin` 词汇(`eval_contract` / `protected_paths`),应能几乎无返工地导出(与 2.2 配套)。 -### 2.2 插件库 +### 2.2 插件库 {#sec-2-2} 在 `mle_kaggle` 之外提供更多范例领域插件,与上面的 Task Pack 配对,让把 Arbor 重定向到 一个领域只需改一行 `plugin:`。 @@ -138,5 +138,44 @@ --- +## 方向四 —— 采用、开发者体验与社区 + +前三个方向扩展 Arbor *能做什么*;这个方向降低"好奇的访客"到"第一次成功运行" +之间的门槛,并让项目的推进势头可见。目标是缩短到达"aha 时刻"的路径,而不是堆叠 +更多功能面。 + +### 4.1 免安装的在线 Demo + +`arbor replay --demo --html` 已经能导出一个自包含、零依赖的真实 run 页面。把它 +发布出去(例如随文档一起部署到 GitHub Pages),并从 README 顶部链接过去,让访客 +**无需安装任何东西**就能看假设树生长。一份录制好的 run,随仪表盘变化刷新即可。 + +### 4.2 示例画廊 + +目前 `examples/algotune_knn` 是新手唯一能端到端跑通的任务。把它扩成一个小型画廊, +让不同受众都能对号入座,每个都配可复制的命令和一段 asciinema 录屏: + +- 一个 Kaggle / MLE 风格的任务(与 `mle_kaggle` 插件配套), +- 一个 prompt / harness 工程任务,以及 +- 一个小规模训练 / 微调任务。 + +门槛保持在"笔记本上几分钟、或用免费 key 即可跑通",复用 `algotune_knn` 的零配置 +纪律。 + +### 4.3 规律发布与更新日志 + +版本号已经通过 setuptools-scm 从 `v*` tag 自动派生,发布成本很低。以可预期的节奏 +发布版本并附上人类可读的发布说明,并维护一份 `CHANGELOG.md`(或以 GitHub Releases +为准),让项目进展从外部看也一目了然。 + +### 4.4 公开路线图与 issue 看板 + +把本文档以公开看板(GitHub Projects)的形式呈现,并给跟踪中的工作打上 +`good first issue` / `help wanted` 标签。有几条线索天然适合外部贡献——把 +[benchmark zoo](#sec-2-1) 扩到 3–5 个 pack([2.1](#sec-2-1))、增加领域插件 +([2.2](#sec-2-2))——把它们开放出来,就能把读者变成贡献者。 + +--- + 有想法,或想认领其中一条线索?开一个 [discussion](https://github.com/RUC-NLPIR/Arbor/discussions),或见 [贡献](contributing.md)。