Skip to content

Arbor

external · status: active · focus: end-to-end · discipline: general · started: 2026

Project page: https://github.com/RUC-NLPIR/Arbor

Source: projects/landscape/arbor.yml

Positioning

A generalist autonomous research agent (arXiv:2606.11926) from the WebThinker group at Renmin University: given a goal, a benchmark directory, and a metric, a Coordinator grows a persistent hypothesis tree ("Idea Tree") while Executors implement each idea in isolated git worktrees, run real experiments on a dev split, and merge only gains that survive a held-out test split. Sits in the experiment-driven optimization corner of the landscape alongside CORAL and AlphaEvolve rather than the paper-writing corner.

Distinctive contribution

The hypothesis tree as first-class research memory: results, failure modes, and distilled insights backpropagate to ancestor nodes so later ideas inherit lessons, and cross-run memory carries findings into future tasks. Ships three ways — native CLI, a Claude Code/Codex markdown skill suite, and a keyless MCP mode where the host harness's model does the reasoning while Arbor supplies deterministic tree/eval/merge/report tools.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Five stages from hypothesis generation through experiment execution; no paper-drafting or review stage — the output is a run report and merged code, not a manuscript.
Autonomy level 3 Fully autonomous in auto mode once the intake Research Contract is confirmed; optional direction/review/collaborative human-in-the-loop modes.
Architectural transparency 3 Apache-2.0 code, arXiv paper, docs site, example configs, and the agent skill suite as plain markdown; orchestration and runtime skills all public.
Inputs supported 2 Goal string, YAML config, or existing benchmark repo; keyless alphaXiv literature access; experimental arbor benchmark add assembles tasks from a one-line request.
Outputs / reproducibility 2 Git-worktree-isolated experiments with session artifacts (REPORT.md, events.jsonl, Idea Tree), resumable runs, and deterministic replay/export; LLM nondeterminism limits exact re-runs.
Internal evaluation 2 Systematic paper evaluation across six tasks plus MLE-Bench Lite (86.36% Any-Medal) against Claude Code and Codex baselines; self-reported, not yet peer-reviewed.
Openness 3 Apache-2.0; pip-installable; keyless demo replay and a CPU-only example task run on commodity hardware, with free-key/local-model quickstart paths.
Maturity / traction 2 968 stars and 118 forks in ~6 weeks, PyPI package, Claude Code plugin marketplace entry, VentureBeat coverage; very young but shipping regular releases.
Cross-family policy 0 Coordinator and Executor share one configured provider/model per run; many back-ends supported (Anthropic, OpenAI, LiteLLM-compatible) but one family at a time.
Runtime assurance 2 Held-out-test merge gates with a configurable margin, protected dev/test discipline, git isolation, and optional pre-experiment novelty checks — moderate, metric-focused gating.
Cross-platform portability 3 Native CLI, Claude Code plugin, Codex skills, standalone markdown skill suite, and an MCP server; model back-ends span Anthropic, OpenAI, and any OpenAI-compatible endpoint via LiteLLM.

Scored on 2026-07-23. See the evaluation rubric.

Tags

Pipeline stages: hypothesis-generation literature-discovery research-design code-generation data-analysis

Architectural features: multi-agent tool-use iterative-loop persistent-memory artifact-versioning

Inputs: research-goal codebase evaluation-script

Outputs: optimized-codebase idea-tree run-report session-event-log

Data sources: user-provided

Knowledge sources: alphaxiv-literature cross-run-memory

Limitations

  • Optimization-shaped: requires a runnable evaluation script and a measurable metric; no paper-drafting or peer-review stages, so it optimizes artifacts rather than authoring scholarship.
  • Headline results (2.5x over Claude Code/Codex on equal compute, MLE-Bench Lite medals) are self-reported by the authors; no third-party replication or peer review yet.
  • Literature grounding is off by default inside runs, and the bundled alphaXiv novelty-check backend requires Python >= 3.12.

Papers describing this project

  • Toward Generalist Autonomous Research via Hypothesis-Tree Refinement — Jin, J., Hu, Y., Qiu, K., Dai, Q., Luo, C., Dong, G., et al. (2026). arXiv. arXiv:2606.11926