Toward Generalist Autonomous Research via Hypothesis-Tree Refinement

Problem
This work addresses the gap in autonomous research capabilities, specifically the lack of frameworks that can effectively manage long-term research processes without human intervention. The authors propose a novel approach to automate the iterative loop of exploration, experimentation, and abstraction, which is critical for scientific progress. This paper is a preprint and has not yet undergone peer review.

Method
The core technical contribution is the Arbor framework, which consists of a long-lived coordinator and short-lived executors. The coordinator oversees the global research strategy, while executors are responsible for implementing and testing individual hypotheses within isolated worktrees. The key innovation is the Hypothesis Tree Refinement (HTR), a persistent structure that organizes hypotheses, artifacts, evidence, and insights over time. As results are generated, Arbor updates the HTR, propagates reusable lessons, refines the search frontier, and incorporates verified improvements. The framework is evaluated in an Autonomous Optimization (AO) setting, where the agent autonomously improves an initial research artifact through iterative experimentation without step-level human supervision.

Results
Arbor was tested across six real research tasks, including model training, harness engineering, and data synthesis. It achieved superior performance, with more than 2.5 times the average relative held-out gain compared to Codex and Claude Code under the same task interface and resource budget. Specifically, on the MLE-Bench Lite benchmark, Arbor attained an impressive 86.36% Any Medal using GPT-5.5, marking the highest result in the comparative analysis.

Limitations
The authors acknowledge that the framework’s performance is contingent on the quality of the initial hypotheses and the effectiveness of the executors. They also note that while Arbor excels in the tested tasks, its generalizability to other domains remains to be validated. Additionally, the reliance on a persistent tree structure may introduce complexity in managing large-scale research endeavors. The paper does not address potential scalability issues or the computational overhead associated with maintaining the HTR.

Why it matters
The implications of this work are significant for the field of autonomous AI research, as it provides a structured approach to managing long-term research processes. By enabling an AI agent to iteratively refine hypotheses and learn from past experiments, Arbor could accelerate scientific discovery and innovation. This framework could serve as a foundation for future research in autonomous systems, particularly in areas requiring extensive experimentation and optimization, as published in arXiv cs.AI.