Hu, Zhengyu, Xiao, Zheyuan, Wang, Yike, Ivison, Hamish, Wang, Haoyu, Yuan, Yige, Shao, Rulin, Li, Yuetai, Lyu, Niqi, Ding, Yicheng, Yang, Shu, Zeng, Zhichen, Zhu, Banghua, Pei, Jiaxin, Ding, Kaize, Dasigi, Pradeep, Roth, Dan, Hajishirzi, Hannaneh, Smith, Noah A., Xiao, Teng, Poovendran, Radha · Nov 9, 2026
Infrathon: The authors test language agents on long-horizon AI infrastructure optimization
Markdown · JSON · Original source
The authors of Infrathon ask whether frontier agents can keep improving real-world AI infrastructure over extended work periods. Their covers diverse AI infrastructure optimization tasks, and they evaluate six models across 150 runs with 24-hour budgets. They report that model rankings vary with the evaluation horizon, and that 135 runs improved further in the second half of their budgets. In their results, shorter evaluations can miss both later gains and shifts in which model performs best. The authors also report a case where Opus 5.5 raised the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%. The comparison suggests agents may find infrastructure solutions that outperform human-designed baselines on a measured task; it does not establish that result across other tasks or deployments. The authors identify task violations, unreliable testing and premature completion as barriers to sustained improvement. The abstract presents Infrathon as a resource for evaluating agents over longer periods and discovering infrastructure solutions.
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: Nov 9, 2026 (2026-11-09).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T09:47:26.649Z. First observed: 2026-11-10T08:39:54.499Z. Retrieved: 2026-11-12T09:46:39.211Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the paper's abstract, which does not detail the tasks, evaluation methods, or how the reported barriers affected individual runs. The reported codec comparison is for one stated task; the abstract does not establish performance across other tasks or real-world deployments.
Open question
How much does each of task violations, unreliable testing and premature completion limit sustained improvement in the runs?
Source-stated
The authors introduce Infrathon to study whether frontier agents can continually improve real-world AI infrastructure over long horizons.
In this work, we study whether frontier agent can continually improve real-world AI infrastructure over long horizons. We introduce Infrathon
Source-stated
The authors evaluate six models in 150 runs, each with a 24-hour budget, totaling over 10,000 GPU-hours.
We evaluate six models across 150 runs with 24-hour budgets, totaling over 10,000 GPU-hours.
Source-stated
Model rankings vary with evaluation horizon, and 135 of 150 runs improved further during the second half of their budgets.
model rankings vary with the evaluation horizon, while 135 of the 150 runs achieve further improvements during the second half of their budgets.
Source-stated
The authors say shorter evaluations can miss both optimization gains and changes in relative model performance.
These findings indicate that shorter evaluations can overlook both optimization gains and changes in relative model performance.
Source-stated
The abstract reports Claude Opus 5.5 increased mean KV-cache compression ratio from 6.82% with a human-designed codec to 19.86%.
Claude Opus 5.5 increases the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%.
Source-stated
The authors identify task violations, unreliable testing and premature completion as barriers to sustained improvement.
Analysis of agent trajectories identifies task violations, unreliable testing, and premature completion as barriers to sustained improvement.
Source and contributor context
Hu, Zhengyu, Xiao, Zheyuan, Wang, Yike, Ivison, Hamish, Wang, Haoyu, Yuan, Yige, Shao, Rulin, Li, Yuetai, Lyu, Niqi, Ding, Yicheng, Yang, Shu, Zeng, Zhichen, Zhu, Banghua, Pei, Jiaxin, Ding, Kaize, Dasigi, Pradeep, Roth, Dan, Hajishirzi, Hannaneh, Smith, Noah A., Xiao, Teng, Poovendran, Radha · 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.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 9f05af90-ea85-4d26-8130-c7c88e9ea2fa.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-12T09:47:26.649Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: c2afaa2a003a8073d02a21d2d8fc64672c4f07b1678895965c9275dc7cbe5211. Source text hash: 61a0f9c9c14bac2746e47a97c7fba725c27eb5f52894ad47c2eb33228fd0329c.
Revision and source-check history
Version 1 · 2026-11-12T09:47:26.649Z
Infrathon: The authors test language agents on long-horizon AI infrastructure optimization
The authors of Infrathon ask whether frontier agents can keep improving real-world AI infrastructure over extended work periods. Their covers diverse AI infrastructure optimization tasks, and they evaluate six models across 150 runs with 24-hour budgets. They report that model rankings vary with the evaluation horizon, and that 135 runs improved further in the second half of their budgets. In their results, shorter evaluations can miss both later gains and shifts in which model performs best. The authors also report a case where Opus 5.5 raised the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%. The comparison suggests agents may find infrastructure solutions that outperform human-designed baselines on a measured task; it does not establish that result across other tasks or deployments. The authors identify task violations, unreliable testing and premature completion as barriers to sustained improvement. The abstract presents Infrathon as a resource for evaluating agents over longer periods and discovering infrastructure solutions.
Original source
cs.AI · Article
Source details
Original source date: Nov 9, 2026 (2026-11-09).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T09:47:26.649Z. First observed: 2026-11-10T08:39:54.499Z. Retrieved: 2026-11-12T09:46:39.211Z.
Limits
The supplied evidence is the paper's abstract, which does not detail the tasks, evaluation methods, or how the reported barriers affected individual runs. The reported codec comparison is for one stated task; the abstract does not establish performance across other tasks or real-world deployments.
Open question
How much does each of task violations, unreliable testing and premature completion limit sustained improvement in the runs?
Source-stated
The authors introduce Infrathon to study whether frontier agents can continually improve real-world AI infrastructure over long horizons.
In this work, we study whether frontier agent can continually improve real-world AI infrastructure over long horizons. We introduce Infrathon
Source-stated
The authors evaluate six models in 150 runs, each with a 24-hour budget, totaling over 10,000 GPU-hours.
We evaluate six models across 150 runs with 24-hour budgets, totaling over 10,000 GPU-hours.
Source-stated
Model rankings vary with evaluation horizon, and 135 of 150 runs improved further during the second half of their budgets.
model rankings vary with the evaluation horizon, while 135 of the 150 runs achieve further improvements during the second half of their budgets.
Source-stated
The authors say shorter evaluations can miss both optimization gains and changes in relative model performance.
These findings indicate that shorter evaluations can overlook both optimization gains and changes in relative model performance.
Source-stated
The abstract reports Claude Opus 5.5 increased mean KV-cache compression ratio from 6.82% with a human-designed codec to 19.86%.
Claude Opus 5.5 increases the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%.
Source-stated
The authors identify task violations, unreliable testing and premature completion as barriers to sustained improvement.
Analysis of agent trajectories identifies task violations, unreliable testing, and premature completion as barriers to sustained improvement.
Source and contributor context
Hu, Zhengyu, Xiao, Zheyuan, Wang, Yike, Ivison, Hamish, Wang, Haoyu, Yuan, Yige, Shao, Rulin, Li, Yuetai, Lyu, Niqi, Ding, Yicheng, Yang, Shu, Zeng, Zhichen, Zhu, Banghua, Pei, Jiaxin, Ding, Kaize, Dasigi, Pradeep, Roth, Dan, Hajishirzi, Hannaneh, Smith, Noah A., Xiao, Teng, Poovendran, Radha · 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.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 9f05af90-ea85-4d26-8130-c7c88e9ea2fa.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-12T09:47:26.649Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: c2afaa2a003a8073d02a21d2d8fc64672c4f07b1678895965c9275dc7cbe5211. Source text hash: 61a0f9c9c14bac2746e47a97c7fba725c27eb5f52894ad47c2eb33228fd0329c.
passed · 2026-11-12T09:47:26.649Z · anthropic/claude-opus-5.5
Publication: 9f05af90-ea85-4d26-8130-c7c88e9ea2fa
attribution — pass: The draft consistently attributes the benchmark, results and interpretations to the paper's authors or abstract. The arXiv authorship and the Claude Opus 5.5 result are presented as reported claims, not as established facts.
We introduce Infrathon, a benchmark of diverse, long-horizon AI infrastructure optimization tasks across different levels.
summary — pass: The title, summary and question match the abstract. The reading leads with what the authors test, then explains why it matters: shorter evaluations can miss later gains and shifts in model rankings. The question about the relative weight of the three named barriers is a genuine open issue. KV-cache compression ratio is left unexplained, but it is a supporting example rather than the central development. A term-library tooltip for KV-cache would help readers.
These findings indicate that shorter evaluations can overlook both optimization gains and changes in relative model performance.
date — pass: The original publication date of 2026-11-09 matches the arXiv v1 submission date. It is not the later feed date.
[Submitted on 9 Nov 2026]
support — pass: All six claims are verbatim-supported. The hedge on the codec result stays within the source and does not overstate it: the result is limited to one measured task and does not establish deployment performance. The draft invents no benefits, costs or stakeholders.
Claude Opus 5.5 increases the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%.
media — pass: The selected media is the verified, direct arXiv abstract page for this paper, which is the original 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 defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference