Prakriya, Neha, Hou, Chaojun, Gong, Zheng, Zhao, Huasha, Qin, Yanyuan, Cai, Chris, Zhao, Xi, Gu, Zhenyu, Barsoum, Emad · Jun 10, 2026
Arbor authors: How Search Policy Shapes Agent Behavior in System Optimization
Markdown · JSON · Original source
The authors present Arbor, a system that gives autonomous agents an explicit optimization tree to explore, backtrack through and deepen with experience. They report performance gains across production inference models, training and recommendation training. In controlled ablations using the same agent, tree and tools, different search policies produced different gains through different kinds of changes: targeted framework fixes, deep kernel rewrites and compounding independent changes. An agent-driven loop that selected its own targets improved less and wrote more code per percentage point of improvement than Arbor’s structured policies. The authors’ central argument is that search structure shapes both the magnitude and the of agent optimizations. That distinction matters in a process they describe as requiring weeks of expert effort per model and hardware platform. The reported results show how policy choices affected the tested optimization campaigns; the abstract does not establish whether the gains transfer beyond those settings.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Jun 10, 2026 (2026-06-10).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-19T07:16:19.414Z. First observed: 2026-11-19T05:04:14.209Z. Retrieved: 2026-11-19T07:04:50.711Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the abstract; it does not include the paper’s detailed experimental setup or results needed to assess the reported comparisons. The abstract reports results on specified workloads but does not establish how consistently they would transfer to other models, hardware platforms or optimization campaigns.
Open question
Across the reported workloads, how consistently do the different search policies produce the distinct kinds of optimization the authors describe?
Source-stated
The authors introduce Arbor, which gives autonomous agents an optimization tree to reason over, backtrack through and deepen with experience, using pluggable search policies and a shared harness.
a system that gives autonomous agents an explicit optimization tree to reason over, backtrack through, and deepen with experience
Source-stated
The authors report gains across six production inference models on two GPU generations, as well as LLM training and recommendation training.
+40% to +193% on six production inference models across two GPU generations, +42.8% on LLM training, and +41% on recommendation training
Source-stated
In controlled ablations with the same agent, tree and tools, the authors report different gains and optimization approaches for bandit, sweep and swarm search policies.
The bandit achieves +30.1% through targeted framework fixes, the sweep +26.4% through deep kernel rewrites, and the swarm +33.4% by compounding independent changes.
Source-stated
The authors report that an agent-driven loop choosing its own targets reached a smaller gain and wrote more code per percentage point of improvement than Arbor’s structured policies.
selects its own targets reaches +14%, writing 4x more code per percentage point of improvement than Arbor's structured search policies
Source-stated
The authors introduce an evaluation framework for autonomous loop-based agents, covering four dimensions across long-running optimization campaigns.
spanning efficiency, grounding, robustness, and depth, designed to characterize agent search behavior across long-running optimization campaigns.
Source and contributor context
Prakriya, Neha, Hou, Chaojun, Gong, Zheng, Zhao, Huasha, Qin, Yanyuan, Cai, Chris, Zhao, Xi, Gu, Zhenyu, Barsoum, Emad · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Nature
Nature publishes research, news, and commentary across the sciences.
Editorial provenance
Version 1. Publication ID: 1cc79b62-61a9-4157-801d-be2d81265d08.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T07:16:19.414Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 99d3499edf8ccc8e133c7e82c68f7bae8983da603476a27416dfe57a73fd2fca. Source text hash: 315b546055536a5ff228f6cf1d4fefe0717d574024c9119ed554fbd54b881d6a.
Revision and source-check history
Version 1 · 2026-11-19T07:16:19.414Z
Arbor authors: How Search Policy Shapes Agent Behavior in System Optimization
The authors present Arbor, a system that gives autonomous agents an explicit optimization tree to explore, backtrack through and deepen with experience. They report performance gains across production inference models, training and recommendation training. In controlled ablations using the same agent, tree and tools, different search policies produced different gains through different kinds of changes: targeted framework fixes, deep kernel rewrites and compounding independent changes. An agent-driven loop that selected its own targets improved less and wrote more code per percentage point of improvement than Arbor’s structured policies. The authors’ central argument is that search structure shapes both the magnitude and the of agent optimizations. That distinction matters in a process they describe as requiring weeks of expert effort per model and hardware platform. The reported results show how policy choices affected the tested optimization campaigns; the abstract does not establish whether the gains transfer beyond those settings.
Original source
cs.AI · Article
Source details
Original source date: Jun 10, 2026 (2026-06-10).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-19T07:16:19.414Z. First observed: 2026-11-19T05:04:14.209Z. Retrieved: 2026-11-19T07:04:50.711Z.
Limits
The supplied evidence is the abstract; it does not include the paper’s detailed experimental setup or results needed to assess the reported comparisons. The abstract reports results on specified workloads but does not establish how consistently they would transfer to other models, hardware platforms or optimization campaigns.
Open question
Across the reported workloads, how consistently do the different search policies produce the distinct kinds of optimization the authors describe?
Source-stated
The authors introduce Arbor, which gives autonomous agents an optimization tree to reason over, backtrack through and deepen with experience, using pluggable search policies and a shared harness.
a system that gives autonomous agents an explicit optimization tree to reason over, backtrack through, and deepen with experience
Source-stated
The authors report gains across six production inference models on two GPU generations, as well as LLM training and recommendation training.
+40% to +193% on six production inference models across two GPU generations, +42.8% on LLM training, and +41% on recommendation training
Source-stated
In controlled ablations with the same agent, tree and tools, the authors report different gains and optimization approaches for bandit, sweep and swarm search policies.
The bandit achieves +30.1% through targeted framework fixes, the sweep +26.4% through deep kernel rewrites, and the swarm +33.4% by compounding independent changes.
Source-stated
The authors report that an agent-driven loop choosing its own targets reached a smaller gain and wrote more code per percentage point of improvement than Arbor’s structured policies.
selects its own targets reaches +14%, writing 4x more code per percentage point of improvement than Arbor's structured search policies
Source-stated
The authors introduce an evaluation framework for autonomous loop-based agents, covering four dimensions across long-running optimization campaigns.
spanning efficiency, grounding, robustness, and depth, designed to characterize agent search behavior across long-running optimization campaigns.
Source and contributor context
Prakriya, Neha, Hou, Chaojun, Gong, Zheng, Zhao, Huasha, Qin, Yanyuan, Cai, Chris, Zhao, Xi, Gu, Zhenyu, Barsoum, Emad · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Nature
Nature publishes research, news, and commentary across the sciences.
Editorial provenance
Version 1. Publication ID: 1cc79b62-61a9-4157-801d-be2d81265d08.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T07:16:19.414Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 99d3499edf8ccc8e133c7e82c68f7bae8983da603476a27416dfe57a73fd2fca. Source text hash: 315b546055536a5ff228f6cf1d4fefe0717d574024c9119ed554fbd54b881d6a.
uncertain · 2026-11-19T07:16:19.414Z · anthropic/claude-opus-5.5
Publication: 1cc79b62-61a9-4157-801d-be2d81265d08
attribution — pass: The reading attributes Arbor and all results to the paper's authors as reported claims, consistent with the arXiv listing. Results are framed as reported claims, not established facts.
Authors:Neha Prakriya, Chaojun Hou, Zheng Gong, Huasha Zhao, Yanyuan Qin, Chris Cai, Xi Zhao, Zhenyu Gu, Emad Barsoum
summary — uncertain: The summary is faithful but hard to follow without specialist background. It never says what is being optimized: making AI models run faster on accelerators such as GPUs. Central terms are unexplained, including search policy, optimization tree, controlled ablations, kernel, inference, bandit, sweep and swarm. It also drops the reported magnitudes (+40% to +193%, +14% vs roughly +26-33%, 4x code), which weakens the source's force. Smallest repair: add one plain sentence on the goal of faster model execution on hardware, plain glosses or tooltips for search policy, ablation and kernel, and the key figures.
Achieving peak performance on modern accelerators requires coordinated optimization across kernels, operator fusion, KV-cache management, quantization, scheduling, and communication overlap
date — pass: originalPublished 2026-06-10 matches the v1 submission date. The v2 revision of 17 Nov 2026 and the November feed date are correctly not used.
[Submitted on 10 Jun 2026 (v1), last revised 17 Nov 2026 (this version, v2)]
support — pass: All five claims match verbatim abstract excerpts. The limitations correctly note that only the abstract is available and that transfer to other settings is not established. The question about consistency across workloads follows from the evidence without asserting an answer.
we find that search structure shapes both the magnitude and the nature of the optimizations
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceNature publishes research, news, and commentary across the sciences.
Explore NaturearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition referenceNature publishes research, news, and commentary across the sciences.
Explore NaturearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv