Claude Skill MartBrowse skillsQuick linesLearn by videoTerminal guideWhat is a Skill?
 
← Back to list

Arbor — Autonomous Optimization via Hypothesis Tree Refinement

Runs a long-horizon experiment loop that improves a code/model/prompt artifact against an evaluator, using a persistent hypothesis tree and a held-out test merge gate to avoid overfitting.

Dev & CodingAdvanced★ 46,333⑂ 4,193AI score 9/10Last updated: Sep 21, 2026

What it does

Implements Hypothesis Tree Refinement (HTR) from the Arbor paper as a Claude Code workflow.

  • Pins down the task tuple first: (M_0 artifact, objective, E_dev evaluator, E_test evaluator).
  • Claude acts as a long-lived coordinator owning a hypothesis tree; short-lived executor subagents each test one hypothesis inside an isolated git worktree.
  • Six-step cycle repeated until budget is spent: Observe → Ideate → Select → Dispatch → Backpropagate → Decide.
  • Falsified branches are pruned with a recorded reason, becoming negative constraints; leaf insights are abstracted into direction-level and global priors.
  • A candidate becomes the new best only if it beats the held-out E_test, so dev-set exploitation gets caught instead of shipped.
  • All bookkeeping runs through scripts/tree.py (init, observe, add-node, set-evidence, propagate, prune, merge) plus reference docs for executor briefs and the final report.

Who it's for

  • ML engineers tuning training recipes, optimizers, or architectures over many trials.
  • Agent/harness builders pushing pass rate or accuracy on a scored loop.
  • Anyone doing MLE-bench / Kaggle-style "improve the submission" work.
  • Prompt or data-pipeline optimization where outputs can be auto-scored.
  • Not for one-shot fixes or evaluator-free ideation.

Example uses

  1. Kick off a run: python scripts/tree.py init --objective "Improve BrowseComp answer accuracy" --dev-eval "python eval.py --split dev --n 50" --test-eval "python eval.py --split test --n 300" --branching 3 --max-depth 2 --budget 12.
  2. Compare sibling hypotheses in parallel: under the depth-1 direction "verification, not retrieval, is the bottleneck", dispatch "aggregate K=5 rollouts by evidence dossier" and "decompose the question into atomic constraints" simultaneously, then abstract the shared lesson upward.
  3. Catch an overfit candidate: a branch that jumps on dev but collapses on test gets prune --reason "search-augmented judge overfits dev questions; no test transfer", and that reason steers later ideation away from the same trap.

· · · Install guide · · ·

Try it now, no install

Paste this into Claude to use the skill without installing anything.

Read the instructions in this file and follow them to help me:
https://raw.githubusercontent.com/K-Dense-AI/scientific-agent-skills/HEAD/skills/arbor/SKILL.md

What I want: (describe your task here)

If Claude can't open the link, open it yourself and paste the contents instead.

↓ If it works for you, download the ZIP below and install it. Then it runs on its own — no pasting each time.

Install in the Claude app (no terminal)
  1. Download the ZIP with the button below.
  2. In Claude, open Settings → Capabilities and turn on 'Code execution and file creation'. (one time)
  3. Go to Customize → Skills → + → 'Upload a skill' and upload the ZIP.
↓ Download ZIP
Install in Claude Code

Let Claude do it — paste this into Claude Code

Install the skill I found on Claude Skill Mart.
Copy the skills/arbor folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/arbor/.
When it's done, tell me in one line what this skill can do.

Install with a command instead

git clone https://github.com/K-Dense-AI/scientific-agent-skills.git && mkdir -p ~/.claude/skills && cp -r scientific-agent-skills/skills/arbor ~/.claude/skills/arbor

⚠ This is a third-party skill. Check the source repository before installing.

  1. Open a terminal.
  2. Clone the skill repository: git clone https://github.com/K-Dense-AI/scientific-agent-skills.git
  3. Create the skills directory if needed: mkdir -p ~/.claude/skills
  4. Copy just this skill: cp -r scientific-agent-skills/skills/arbor ~/.claude/skills/arbor
  5. Verify ~/.claude/skills/arbor contains SKILL.md, scripts/tree.py, and the references/ folder.
  6. Make sure Python 3 and git are installed, and that the project you want to optimize is git-tracked and currently runnable.
  7. Prepare two scoring commands: a fast dev evaluator for search, and a separate held-out test evaluator used only at the merge gate.
  8. Restart Claude Code and ask something like "run many experiments to raise this benchmark score without overfitting the dev split" to trigger the skill.
View source on GitHub ↗License: MIT