Youwei Bao, Chengpeng Wang and Simin Chen submitted a preprint on 27 October 2026 proposing batch-invariance checks for GPU kernels generated by agents. The authors report 19 confirmed violations among 83 released artifacts applicable to batching. These are findings reported in the saved arXiv abstract, rather than independently reproduced results.[1]
Large language model (LLM) serving systems dynamically group user requests into batches for GPU execution. Batch invariance means an unchanged input retains the same numerical output when its surrounding batch changes. According to the abstract, kernels that lack this property can produce different numerical results for the same request, undermining reproducibility.[1]
The proposed plug-in combines contract-guided static dependency analysis with dynamic checks of bitwise consistency across batch contexts. Successive acceptance gates reject structurally risky or dynamically non-invariant candidates before performance benchmarking and return feedback to the generation agent. The saved abstract renders the component’s name as an unresolved \tool command.[1]
Across three agent frameworks, the authors report an increase in the batch-invariance pass rate from 67.6% to 91.9%, while retaining 91.8% of the baseline geometric mean speedup. That reported pass rate leaves failures; the abstract does not establish universal batch invariance or identify a fixed production version operators should install.[1]
[1]: https://arxiv.org/abs/2610.36623
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
364576e84a53fc10d47a9f39bfd4ad6a5a0d8cde7d3d52d8a33a551e2090e0ac