Paper 2606.11926
Toward Generalist Autonomous Research via Hypothesis-Tree Refinement
- Published
- Jun 2026
- Research lab
- Independent
- Citations
- 2
- GitHub
- 993 stars
01 In brief
Summary
The paper introduces Arbor, a general framework for autonomous research under the Autonomous Optimization (AO) setting, where an agent iteratively improves an initial artifact without step-level human supervision.
Arbor uses Hypothesis Tree Refinement (HTR), combining a long-lived coordinator that maintains a persistent hypothesis tree (linking hypotheses, artifacts, evidence, and insights) with short-lived executors that test individual hypotheses in isolated worktrees.
A held-out merge gate admits only verified improvements.
Evaluated on six real research tasks (model training, harness engineering, data synthesis), Arbor achieves the best held-out results on all six, with over 2.5x the average relative gain of Codex and Claude Code under the same budget.
On MLE-Bench Lite, Arbor reaches 86.36% Any Medal with GPT-5.5, the strongest result in comparison.
Ablations show the tree and insight feedback are complementary, and a transfer study shows a BrowseComp-optimized harness improves unseen tasks.
Arbor uses 20-43M tokens, comparable to baselines, indicating gains come from structured search rather than larger sampling.
The framework is model-agnostic, with performance depending on task-backbone fit.
Limitations include fixed scalar objectives, evaluation scope, and reliance on model capabilities for idea generation.
Future work includes multi-objective search, broader benchmarks, and cost-aware policies.
The code is open-sourced at https://github.com/RUC-NLPIR/Arbor.
The paper is a living technical report, with ongoing evaluation expansion.
Key contributions include the AO formulation, the Arbor framework with HTR, and a six-task suite demonstrating strong held-out gains.
The system's design separates global strategy from local execution, enabling cumulative research progress through persistent state management.
The results support that persistent…
02 From the paper
Abstract
Scientific progress depends on a repeated loop of exploration, experimentation, and abstraction. Researchers test candidate directions, interpret the evidence, and carry the resulting lessons into later attempts. We study how an AI agent can run this loop autonomously over long horizons. We introduce Arbor, a general framework for autonomous research that combines a long-lived coordinator, short-lived executors, and Hypothesis Tree Refinement (HTR), a persistent tree that links hypotheses, artifacts, evidence, and distilled insights across time. The coordinator manages global research strategy over the tree, while executors implement and test individual hypotheses in isolated worktrees. As results return, Arbor updates the tree, propagates reusable lessons, refines the search frontier, and admits verified improvements. This design turns autonomous research from a sequence of local attempts into a cumulative process in which strategy, execution, and evidence are carried across time. We evaluate Arbor under Autonomous Optimization (AO), an operational setting where an agent improves an initial research artifact through iterative experimentation without step-level human supervision. Across six real research tasks in model training, harness engineering, and data synthesis, Arbor achieves the best held-out result on all six tasks, attaining more than 2.5x the average relative held-out gain of Codex and Claude Code under the same task interface and resource budget. On MLE-Bench Lite, Arbor reaches 86.36% Any Medal with GPT-5.5, the strongest result in our comparison.