Cai, Linrong, Lu, Taiming, Chen, Haozhe, Liu, Zhuang · Nov 5, 2026
XStack tests agent migration from NVIDIA GPUs to Google TPUs
Markdown · JSON · Original source
Linrong Cai, Taiming Lu, Haozhe Chen and Zhuang Liu introduce XStack, a for testing whether agents can migrate repositories from PyTorch on to JAX on Google . The authors frame migration as both a practical need, because it can let researchers use different hardware, and a demanding test: it spans frameworks and hardware, covers whole repositories and takes sustained work. Each agent starts with a PyTorch repository and a clean TPU machine, then must deliver a JAX repository. A result passes when it meets grading criteria based on the source repository’s published result, which the authors say they authenticate and calibrate. They report that Opus 5 passes 75%, GPT-5.6-sol 45%, and every other agent at most 40%. The authors characterize agent migration as unreliable and present XStack as a way to measure progress toward reliable migration across frameworks and hardware. These pass rates describe performance under this benchmark’s grading criteria; the abstract does not establish how well they predict successful migration in other 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: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T23:48:29.956Z. First observed: 2026-11-06T05:03:39.005Z. Retrieved: 2026-11-08T23:47:39.297Z.
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 contains the abstract, not the paper’s full methods or results; it does not give the number or range of repositories used. The reported pass rates concern XStack’s grading criteria. The abstract does not establish how well they predict migration success outside the .
Open question
How do the reported pass rates vary across the repositories in XStack, and what do those results show about the ’s grading criteria?
Source-stated
The authors present XStack as a benchmark for agents migrating machine learning repositories from PyTorch to JAX across different hardware.
We introduce XStack, a benchmark that evaluates agents on migrating ML repositories between these two ecosystems.
Source-stated
The benchmark gives an agent a PyTorch repository and a clean TPU machine, and requires a JAX repository as output.
Each agent receives a PyTorch repository and a clean TPU machine and must deliver a JAX repository.
Source-stated
The authors report pass rates of 75% for Claude Opus 5, 45% for GPT-5.6-sol, and at most 40% for every other agent.
Claude Opus 5 passes 75%, GPT-5.6-sol 45%, and every other agent at most 40%.
Source-stated
The authors say the benchmark’s grading criteria use the source repository’s published result, authenticated and calibrated by the authors.
based on the source repository's published result, authenticated and calibrated by us.
Source-stated
The authors describe migration by agents as unreliable and say XStack is intended to make progress toward reliable migration measurable.
Migration by agents remains unreliable: Claude Opus 5 passes 75%
Source and contributor context
Cai, Linrong, Lu, Taiming, Chen, Haozhe, Liu, Zhuang · 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.
Graphics processing unit
A chip that performs many calculations at once; the main hardware used to train and run AI models.
Tensor processing unit
A chip designed by Google for the calculations used in training and running AI models.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Machine learning
Methods that let computers learn patterns from examples instead of following rules written by hand.
NVIDIA
NVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
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: 0b43b580-3908-49b7-8966-4ccb6956ea6a.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T23:48:29.956Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4f1130dd2c3f9750497cf2e6d047443c2536c4db30b9da8e76c7e79903658e8c. Source text hash: ae5581735bc4fbd15bbef0fd523ec468e6d4ad67be1264ae767b5ea45258a9c0.
Revision and source-check history
Version 1 · 2026-11-08T23:48:29.956Z
XStack tests agent migration from NVIDIA GPUs to Google TPUs
Linrong Cai, Taiming Lu, Haozhe Chen and Zhuang Liu introduce XStack, a for testing whether agents can migrate repositories from PyTorch on to JAX on Google . The authors frame migration as both a practical need, because it can let researchers use different hardware, and a demanding test: it spans frameworks and hardware, covers whole repositories and takes sustained work. Each agent starts with a PyTorch repository and a clean TPU machine, then must deliver a JAX repository. A result passes when it meets grading criteria based on the source repository’s published result, which the authors say they authenticate and calibrate. They report that Opus 5 passes 75%, GPT-5.6-sol 45%, and every other agent at most 40%. The authors characterize agent migration as unreliable and present XStack as a way to measure progress toward reliable migration across frameworks and hardware. These pass rates describe performance under this benchmark’s grading criteria; the abstract does not establish how well they predict successful migration in other settings.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T23:48:29.956Z. First observed: 2026-11-06T05:03:39.005Z. Retrieved: 2026-11-08T23:47:39.297Z.
Limits
The supplied evidence contains the abstract, not the paper’s full methods or results; it does not give the number or range of repositories used. The reported pass rates concern XStack’s grading criteria. The abstract does not establish how well they predict migration success outside the .
Open question
How do the reported pass rates vary across the repositories in XStack, and what do those results show about the ’s grading criteria?
Source-stated
The authors present XStack as a benchmark for agents migrating machine learning repositories from PyTorch to JAX across different hardware.
We introduce XStack, a benchmark that evaluates agents on migrating ML repositories between these two ecosystems.
Source-stated
The benchmark gives an agent a PyTorch repository and a clean TPU machine, and requires a JAX repository as output.
Each agent receives a PyTorch repository and a clean TPU machine and must deliver a JAX repository.
Source-stated
The authors report pass rates of 75% for Claude Opus 5, 45% for GPT-5.6-sol, and at most 40% for every other agent.
Claude Opus 5 passes 75%, GPT-5.6-sol 45%, and every other agent at most 40%.
Source-stated
The authors say the benchmark’s grading criteria use the source repository’s published result, authenticated and calibrated by the authors.
based on the source repository's published result, authenticated and calibrated by us.
Source-stated
The authors describe migration by agents as unreliable and say XStack is intended to make progress toward reliable migration measurable.
Migration by agents remains unreliable: Claude Opus 5 passes 75%
Source and contributor context
Cai, Linrong, Lu, Taiming, Chen, Haozhe, Liu, Zhuang · 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.
Graphics processing unit
A chip that performs many calculations at once; the main hardware used to train and run AI models.
Tensor processing unit
A chip designed by Google for the calculations used in training and running AI models.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Machine learning
Methods that let computers learn patterns from examples instead of following rules written by hand.
NVIDIA
NVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
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: 0b43b580-3908-49b7-8966-4ccb6956ea6a.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T23:48:29.956Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4f1130dd2c3f9750497cf2e6d047443c2536c4db30b9da8e76c7e79903658e8c. Source text hash: ae5581735bc4fbd15bbef0fd523ec468e6d4ad67be1264ae767b5ea45258a9c0.
uncertain · 2026-11-08T23:48:29.956Z · anthropic/claude-opus-5.5
Publication: 0b43b580-3908-49b7-8966-4ccb6956ea6a
attribution — pass: The four named authors match the arXiv listing. The framing, grading method, pass rates and unreliability judgment are all attributed to the authors, and the closing caveat is presented as a limit of the abstract rather than as a source claim.
Authors:Linrong Cai, Taiming Lu, Haozhe Chen, Zhuang Liu
summary — uncertain: The reading is accurate overall, with two fidelity gaps. First, the source term long-horizon is paraphrased as 'takes sustained work'. Repair: keep 'long-horizon' and explain it with a tooltip. Second, the authors' thesis is softened: they argue that ML code should run where researchers choose and call XStack a step toward freeing ML code from its ecosystem. Repair: add one sentence restoring that stance. The central terms PyTorch, JAX, GPU, TPU and agent need term-library tooltips so readers without a specialist background can follow. The question is substantive and not leading.
ML code should run where researchers choose, not where it was first written.
date — pass: originalPublished is 2026-11-05, which matches the arXiv v1 submission date. The feed date of Nov 6 is a later announcement, so it correctly does not replace the original publication date.
[v1] Thu, 5 Nov 2026 18:55:54 UTC (367 KB)
support — pass: All claims match the abstract verbatim, including the task setup (PyTorch repository to JAX on a clean TPU machine), the grading basis and the pass rates. The statement that migration can let researchers use different hardware tracks 'unlocks different hardware'. The final caveat correctly avoids turning benchmark scores into real-world reliability. No unsupported benefits or costs are implied.
a practical need, since it unlocks different hardware for researchers
media — pass: The selected media is the verified arXiv abstract page, which is the direct original source. No better source media was found.
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 referenceMethods that let computers learn patterns from examples instead of following rules written by hand.
Definition referenceNVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Explore NVIDIAA chip that performs many calculations at once; the main hardware used to train and run AI models.
Definition referenceA chip designed by Google for the calculations used in training and running AI models.
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 referenceMethods that let computers learn patterns from examples instead of following rules written by hand.
Definition referenceNVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Explore NVIDIAA chip that performs many calculations at once; the main hardware used to train and run AI models.
Definition referenceA chip designed by Google for the calculations used in training and running AI models.
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