RUC-NLPIR/Arbor

A generalist autonomous research agent — runs experiments, researches, and iteratively optimizes, autonomously.

What it solves

Arbor is an autonomous research agent designed for general-purpose optimization. It solves the problem of "forgetting" failed attempts during long-horizon exploration by maintaining a structured hypothesis tree. This allows it to optimize any task that has a measurable target and metric—such as model training, data synthesis, or harness engineering—without wasting compute on redundant or failed ideas.

How it works

Arbor employs two cooperating agents: a Coordinator (the research director) and an Executor (the research engineer).

They operate in a six-step "arbor cycle":

  1. Observe: Analyze current results and failure modes.
  2. Ideate: Propose new hypotheses based on the Idea Tree.
  3. Select: Pick the most promising ideas to test.
  4. Dispatch: The Executor implements the idea in an isolated git worktree and evaluates it on a development split.
  5. Backpropagate: Record results and abstract insights upward through the tree.
  6. Decide: Use a held-out test split to decide whether to merge the gain into the main trunk or prune the branch.

It can be used as a standalone CLI, integrated into coding agents like Claude Code or Codex via MCP, or as a standalone skill suite.

Who it’s for

Arbor is built for researchers and developers who need to iteratively optimize code or models against a specific benchmark and want a disciplined, automated way to explore hypotheses while avoiding overfitting to development data.

Highlights

  • Hypothesis Tree: Maintains a structured history of insights and failures to guide future exploration.
  • Real Experiment Discipline: Uses isolated git worktrees and held-out validation sets to ensure only genuine gains are merged.
  • Literature Grounding: Integrates with the alphaXiv API to check the novelty of ideas before spending compute.
  • Flexible Deployment: Supports multiple LLM backends (Anthropic, OpenAI, DeepSeek, etc.) and integrates directly into existing AI coding harnesses.
  • Steerable: Offers a live dashboard and multiple interaction modes (auto, direction, review, collaborative) for human-in-the-loop guidance.