Daily Paper Briefing

Abstract Overview

This paper studies autonomous research as a long-horizon optimization problem and formalizes it as Autonomous Optimization (AO), where an agent must iteratively improve an artifact using development feedback while reserving held-out evaluation for admission decisions. The proposed system, Arbor, combines a long-lived coordinator with short-lived executors and organizes the search process through Hypothesis Tree Refinement (HTR), a persistent tree linking hypotheses, artifact versions, evidence, and distilled insights. In this design, executors test individual hypotheses in isolated worktrees, while the coordinator updates the tree, propagates lessons, refines the search frontier, and merges candidates only when they improve held-out performance. The evaluation covers six real research tasks spanning model training, harness engineering, and data synthesis, plus MLE-Bench Lite, to assess both task generality and long-horizon ML engineering performance.

Novelty

The distinctive contribution is the use of a persistent hypothesis tree as the operational research state for an autonomous agent, rather than treating experimentation as a flat sequence of local trials. Arbor further pairs this structure with a coordinator-executor split and a held-out merge gate, so that reusable insights and verified improvements accumulate across time in an auditable way.

Results

Across six real AO tasks, Arbor achieves the best held-out result on all six and is reported to deliver more than 2.5× the average relative held-out gain of Codex and Claude Code under the same interface and resource budget. On MLE-Bench Lite, Arbor reaches 86.36% Any Medal with GPT-5.5, which the paper reports as the strongest result in its comparison. Ablations indicate that both the tree structure and insight propagation contribute materially, with performance dropping when either is removed.

Key Points

1. Arbor represents research progress as a persistent hypothesis tree that stores hypotheses, implementation references, evaluation evidence, and distilled insights across iterations.
2. The framework separates global strategy from local execution: a long-lived coordinator manages frontier expansion, pruning, and held-out admission, while short-lived executors run hypothesis-bound experiments in isolated worktrees.
3. Empirically, the method shows broad gains across model training, harness engineering, and data synthesis tasks, and the paper argues that the improvements come from structured evidence accumulation rather than simply using a larger token budget.

References

• arXiv: https://arxiv.org/abs/2606.11926v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2606.11926v1
• Hugging Face Papers: https://huggingface.co/papers/2606.11926