Hypothesis-Tree Refinement: Structured Memory for Research Agents

The provided text introduces *Arbor**, a framework designed for **Autonomous Optimisation (AO)* that enables AI agents to conduct long-term scientific research without constant human guidance. It addresses the limitations of current agents by using a *Hypothesis-Tree Refinement (HTR)* system, which organises experiments into a persistent tree structure of ideas, evidence, and insights. This method separates a long-lived *coordinator**, who manages global strategy, from short-lived **executors* who implement specific code changes in isolated environments. By maintaining this structured memory, *Arbor* turns a sequence of disconnected trials into a cumulative process where failed attempts inform future success. Evaluation across diverse tasks, including *model training* and *data synthesis**, shows that the system significantly outperforms existing tools like Codex and Claude Code. Ultimately, the research demonstrates that **externalising the research state* through a hierarchical tree allows AI to achieve more reliable and sophisticated scientific progress. *Title:* Toward Generalist Autonomous Research via Hypothesis-Tree Refinement *Authors and Institutions:* Jiajie Jin, Yuyang Hu, Guanting Dong, Xiaoxi Li, Tong Zhao, Hongjin Qian, Yutao Zhu, and Zhicheng Dou (Gaoling School of Artificial Intelligence, Renmin University of China); Kai Qiu, Qi Dai, Chong Luo, Xiaolong Ma, Gongrui Zhang, Zhirong Wu, Bei Liu, Zhengyuan Yang, Linjie Li, and Lijuan Wang (Microsoft Research). *What problem the paper was trying to solve* The paper addresses the challenge of **Autonomous Optimization (AO)**, where AI agents must autonomously conduct long-horizon research without getting lost in a flat sequence of local trial-and-error attempts. Because scientific research involves delayed feedback, costly experiments, and frequent failures, agents typically struggle to maintain a coherent research state that remembers past attempts and lets lessons reshape future exploration without step-level human supervision. *What are the paper's key novel ideas?* The central novel idea is **Hypothesis Tree Refinement (HTR)**, which serves as a persistent, structured memory of the research process. Instead of treating experiments as isolated tool calls, HTR organizes them into a tree where each node binds together a specific hypothesis, the code artifact that realizes it, the experimental evidence, and the distilled insights. This transforms transient, local failures and successes into cumulative, reusable constraints that actively guide future ideation. *What is the architecture or method they are using?* The researchers introduce a framework called *Arbor**, which splits research tasks between a **long-lived coordinator* and *short-lived executors**. The coordinator manages the global hypothesis tree—observing the frontier, proposing refinements, and abstracting insights upward to update its global understanding. The executors are strictly constrained to test single hypotheses in isolated workspaces and return structured evidence (scores, results, and insights). Finally, Arbor utilizes a **held-out merge gate* that ensures an artifact is only updated if the new hypothesis demonstrably improves performance on a hidden test set, preventing the agent from overfitting to development data. *Why the paper matters* This research represents a major step toward making AI agents capable of sustained, self-directed scientific progress. By imposing a structured research state, Arbor achieved the *best held-out results across six distinct, real-world research tasks**, attaining more than 2.5 times the average relative gain of strong baseline agents like Codex and Claude Code. It also achieved an **86.36% "Any Medal" rate on the MLE-Bench Lite* benchmark, proving that structured hypothesis management consistently outperforms flat experimental queues. *What are the potential applications* The Arbor framework can be broadly applied to autonomously discover and refine complex software and machine learning systems. The paper demonstrates its direct applicability in *model training* (designing more efficient neural network optimizers and architectures), *harness engineering* (improving the evaluation and control logic around other AI agents), and *data synthesis* (generating better synthetic data pipelines for model training and reasoning evaluation). The description, research summary based on a human-derived template and the video were generated by Google's NotebookLM on 7th July 2026.