Arbor paper: search policy shapes how autonomous agents optimize AI systems
An arXiv paper describes Arbor, a system where autonomous agents optimize model performance using an explicit search tree, with the choice of search policy changing both results and methods.
In plain words
Tuning AI systems for speed on specific hardware normally takes expert engineers weeks. Arbor lets autonomous agents explore an organized tree of optimization choices, backtracking and building on past attempts. The paper reports that which search strategy the agent follows changes both how large the gains are and what kind of code changes the agents produce.
What was announced
The authors, presenting Arbor on arXiv, say peak performance on modern accelerators requires coordinated work across kernels, operator fusion, KV-cache management, quantization, scheduling, and communication overlap, typically demanding weeks of expert effort per model and hardware platform. Arbor gives agents an explicit optimization tree to reason over, backtrack through, and deepen with experience, with pluggable search policies, bandit traversal, parallel agent swarms, and depth-first sweeps, all sharing a harness of containerized isolation, automatic rollback, and end-to-end validation.
Reported results include gains of 40% to 193% on six production inference models across two GPU generations, 42.8% on LLM training, and 41% on recommendation training, with agents writing kernel- and framework-level code spanning thousands of lines. In ablations giving every policy the same agent, tree, and tools, the bandit reached 30.1% via targeted framework fixes, the sweep 26.4% via deep kernel rewrites, and the swarm 33.4% by compounding independent changes. An agent-driven loop with the same harness but free target selection reached 14%, writing four times more code per percentage point of improvement. These are the authors' own measurements, not independently confirmed.
What it does not show
- The source is the paper's abstract only; full experimental setup, baselines, and reproducibility details are not visible in the excerpt.
- All performance figures are reported by the paper's authors; no independent verification is provided.
For security and IT teams
- Scope
- The paper also introduces an evaluation framework for loop-based autonomous agents covering efficiency, grounding, robustness, and depth in long-running optimization campaigns.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.