Bao, Youwei, Wang, Chengpeng, Chen, Simin · Oct 27, 2026
God Does Not Play Dice: Enforcing Batch Invariance in Agentic Kernel Generation
Markdown · JSON · Original source
Youwei Bao, Chengpeng Wang and Simin Chen argue that kernel-generation agents should check batch invariance as well as correctness and speed. In serving systems, a request may run alongside different requests in different batches. Batch invariance (BI) means that changing this batch context does not change the numerical output for an unchanged input; without it, results may be less reproducible. The authors report 19 confirmed BI violations among 83 batch-applicable released artifacts. They propose a plug-in component for kernel agents that combines contract-guided static dependency analysis with dynamic checks for bitwise consistency across batch contexts. In their approach, successive acceptance gates reject candidates judged structurally risky or dynamically non-invariant before performance , and return feedback to the agent. Across three agent frameworks on real-world LLM serving tasks, the authors report that their method raised the BI pass rate from 67.6% to 91.9%, while retaining 91.8% of the baseline geometric mean speedup. These are reported experimental results, not evidence here of performance across other systems or uses. How well would the approach transfer to kernel agents and serving tasks beyond those evaluated?
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: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T05:16:35.770Z. First observed: 2026-10-28T08:42:33.197Z. Retrieved: 2026-10-30T05:17:19.723Z.
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 evidence is an abstract; it does not provide the experimental details needed to assess the evaluation or its results independently. The reported experiments cover three agent frameworks and real-world serving tasks; the evidence does not establish how the approach performs in other settings. The abstract reports pass-rate and speedup results, but does not establish that the method improves reliability in deployed systems.
Open question
How well would the approach transfer to kernel agents and serving tasks beyond those evaluated?
Source-stated
The authors say dynamically batched LLM serving can change the numerical output for an unchanged input when GPU kernels lack batch invariance, affecting reproducibility.
this changing batch context can alter the numerical output for an unchanged input, undermining reproducibility.
Source-stated
The authors report 19 confirmed batch-invariance violations among 83 batch-applicable released artifacts.
We find 19 confirmed BI violations among 83 batch-applicable released artifacts.
Source-stated
The proposed component combines contract-guided static dependency analysis with dynamic validation of bitwise consistency across batch contexts.
combines contract-guided static dependency analysis with dynamic validation of bitwise consistency across batch contexts.
Source-stated
The authors report that their method raises the BI pass rate from 67.6% to 91.9% while retaining 91.8% of baseline geometric mean speedup in experiments across three agent frameworks on real-world LLM serving tasks.
improves the BI pass rate from 67.6\% to 91.9\% while retaining 91.8\% of the baseline geometric mean speedup.
Source-stated
The authors position batch-invariance enforcement as an additional reliability objective alongside correctness and speed.
as an additional reliability objective alongside correctness and speed.
Source and contributor context
Bao, Youwei, Wang, Chengpeng, Chen, Simin · 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.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Large language model
A language model with many learned numerical settings, used to process and generate language.
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.
Editorial provenance
Version 1. Publication ID: fff28868-1ea5-4500-8044-9e1a5766b10c.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T05:16:35.770Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: e9e64c09d6d9482ce83cc6264af9061e5e83ea0ef86961f70a505358b50250bc. Source text hash: 392d7b74503e17ad15b50de58b6feb588c48c0bcf2560c2119ceb04dbf0440e1.
Revision and source-check history
Version 1 · 2026-11-14T05:16:35.770Z
God Does Not Play Dice: Enforcing Batch Invariance in Agentic Kernel Generation
Youwei Bao, Chengpeng Wang and Simin Chen argue that kernel-generation agents should check batch invariance as well as correctness and speed. In serving systems, a request may run alongside different requests in different batches. Batch invariance (BI) means that changing this batch context does not change the numerical output for an unchanged input; without it, results may be less reproducible. The authors report 19 confirmed BI violations among 83 batch-applicable released artifacts. They propose a plug-in component for kernel agents that combines contract-guided static dependency analysis with dynamic checks for bitwise consistency across batch contexts. In their approach, successive acceptance gates reject candidates judged structurally risky or dynamically non-invariant before performance , and return feedback to the agent. Across three agent frameworks on real-world LLM serving tasks, the authors report that their method raised the BI pass rate from 67.6% to 91.9%, while retaining 91.8% of the baseline geometric mean speedup. These are reported experimental results, not evidence here of performance across other systems or uses. How well would the approach transfer to kernel agents and serving tasks beyond those evaluated?
Original source
cs.AI · Article
Source details
Original source date: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T05:16:35.770Z. First observed: 2026-10-28T08:42:33.197Z. Retrieved: 2026-10-30T05:17:19.723Z.
Limits
The evidence is an abstract; it does not provide the experimental details needed to assess the evaluation or its results independently. The reported experiments cover three agent frameworks and real-world serving tasks; the evidence does not establish how the approach performs in other settings. The abstract reports pass-rate and speedup results, but does not establish that the method improves reliability in deployed systems.
Open question
How well would the approach transfer to kernel agents and serving tasks beyond those evaluated?
Source-stated
The authors say dynamically batched LLM serving can change the numerical output for an unchanged input when GPU kernels lack batch invariance, affecting reproducibility.
this changing batch context can alter the numerical output for an unchanged input, undermining reproducibility.
Source-stated
The authors report 19 confirmed batch-invariance violations among 83 batch-applicable released artifacts.
We find 19 confirmed BI violations among 83 batch-applicable released artifacts.
Source-stated
The proposed component combines contract-guided static dependency analysis with dynamic validation of bitwise consistency across batch contexts.
combines contract-guided static dependency analysis with dynamic validation of bitwise consistency across batch contexts.
Source-stated
The authors report that their method raises the BI pass rate from 67.6% to 91.9% while retaining 91.8% of baseline geometric mean speedup in experiments across three agent frameworks on real-world LLM serving tasks.
improves the BI pass rate from 67.6\% to 91.9\% while retaining 91.8\% of the baseline geometric mean speedup.
Source-stated
The authors position batch-invariance enforcement as an additional reliability objective alongside correctness and speed.
as an additional reliability objective alongside correctness and speed.
Source and contributor context
Bao, Youwei, Wang, Chengpeng, Chen, Simin · 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.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Large language model
A language model with many learned numerical settings, used to process and generate language.
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.
Editorial provenance
Version 1. Publication ID: fff28868-1ea5-4500-8044-9e1a5766b10c.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T05:16:35.770Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: e9e64c09d6d9482ce83cc6264af9061e5e83ea0ef86961f70a505358b50250bc. Source text hash: 392d7b74503e17ad15b50de58b6feb588c48c0bcf2560c2119ceb04dbf0440e1.
uncertain · 2026-11-14T05:16:35.770Z · anthropic/claude-opus-5.5
Publication: fff28868-1ea5-4500-8044-9e1a5766b10c
attribution — pass: The reading correctly credits the paper to Youwei Bao, Chengpeng Wang and Simin Chen on arXiv. The violation counts, the proposed method and the experimental results all stay framed as the authors' findings or reports.
Authors:Youwei Bao, Chengpeng Wang, Simin Chen
summary — uncertain: The core argument is faithful, and batch invariance is explained well. However, the first sentence centres on 'kernel-generation agents' and later says 'kernel agents', but never says these are GPU kernels. The source's link to GPU execution is dropped. A non-specialist could not tell what a kernel is, and might read it as an operating-system kernel. Smallest repair: describe them as agents that generate GPU kernels, meaning low-level code that runs model computations on GPUs. Add a term-library tooltip for 'GPU kernel'. Keep the other technical terms exact, with tooltips.
If the GPU kernels lack batch invariance (BI), this changing batch context can alter the numerical output for an unchanged input
date — pass: originalPublished is 2026-10-27, which matches the arXiv v1 submission date. The later feed date of 2026-10-29 is correctly not used.
[Submitted on 27 Oct 2026]
support — pass: Every claim has an exact supporting excerpt from the abstract. The 67.6% to 91.9% pass rate and the 91.8% speedup retention are labelled as reported results. The limitations correctly say these do not show deployed reliability. The closing question follows from the stated scope of three frameworks and does not imply benefits or costs that the source does not support.
Experiments across three agent frameworks under baseline and BI-aware generation on real world LLM serving tasks
media — pass: The selected media is the verified direct link to the arXiv abstract page, which is the original source. It is the most direct and relevant way to reach the paper.
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 referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv 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 referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv 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 reference