Arbor builds a persistent hypothesis tree for AI researchers

Ask any current AI coding agent to fix a bug, and it does fine. Ask it to run an open-ended research project over three days, and it inevitably falls apart. It tries an idea, hits a wall, forgets what failed, and starts running in circles.

Basically, frontier agents have had a continuity problem. They are great at short sprints, but terrible at running a multi-week marathon.

Researchers published a paper introducing Arbor, an autonomous research framework designed to fix this amnesia by giving AI a living, persistent "research tree."

A long-lived boss with disposable interns

Instead of cramming hundreds of failed command logs into a giant context window, Arbor introduces a mechanism called Hypothesis-Tree Refinement (HTR).

The framework splits the job between two distinct roles. At the top sits a long-lived coordinator that manages overall research strategy. It looks at the big picture, decides which hypotheses are worth exploring, and keeps track of what the system has learned so far.

Below it are short-lived executors. When the coordinator wants to test an idea, it spins up an executor inside an isolated worktree. The executor writes the code, runs the test, logs the evidence, and then gets decommissioned.

Every node on Arbor's tree tracks the initial hypothesis, the artifacts, raw evidence, and distilled insights. Crucially, failed attempts aren't discarded—their lessons are factored back into the tree to prune dead ends.

Leaving Claude Code and Codex behind

This structural separation paid off heavily when tested against real engineering and machine learning problems.

Under an Autonomous Optimization setup requiring no step-level human intervention, Arbor tackled six real research tasks spanning model training, test harness engineering, and data synthesis. Across all six tasks, Arbor delivered the best held-out performance, notching over 2.5x the average relative gains of Claude Code and Codex under identical resource limits.

When evaluated on the competitive MLE-Bench Lite suite paired with GPT-5.5, Arbor hit an 86.36% Any Medal rate, setting the top mark in the comparison.

Why it matters

For machine learning engineers, Arbor provides a working blueprint for true autonomous research. Instead of babysitting an agent through terminal prompts every ten minutes, you could soon hand a system an open research objective and let it iterate reliably in the background.

We think Arbor’s cleverest design choice is treating execution as ephemeral and memory as a structured tree. Dumping execution noise out of the master context keeps the strategist sharp and prevents catastrophic forgetting.

The obvious catch right now is compute. Spawning multiple isolated worktrees and running iterative trials will rack up hefty inference and infrastructure bills. But by turning chaotic trial-and-error into a cumulative science loop, Arbor shifts AI agents from reactive coders to genuine autonomous investigators.