AI Aulis

AI NEWS IN ENGLISH · 22 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

21 stories · next issue tomorrow morning · How the paper is made →

← Issue 22 November 2026Paper

TensorLean proved 83 of 149 kernel instances correct against their specifications, researchers say

AI-generated contentWritten by AI

github.com, marktechpost.com, panewslab.com, u.today, arXiv · 5 sources · Sources published: 16 November 2025–22 November 2026 · Some source publication dates are unknown

TensorLean, a Lean-based tool from Columbia University and CertiK, found 38–39 bugs in the other 66, including in FlashAttention and DeepSeek’s DeepGEMM. The tool and the paper have not yet been released.

GPU kernels are programs that carry out the calculations for AI model training and inference on a GPU. TensorLean does not run them on test inputs. Instead, it proves in the Lean proof assistant that the PTX code a compiler produces, the low-level instruction language of NVIDIA GPUs, matches the tensor-level specification the developer wrote. CertiK announced the results on 20 November.

The project’s GitHub page says the proof covers three things. The code has no data races, in which parallel threads access the same memory without synchronisation. The code also terminates and computes as the specification requires, within floating-point tolerance. The proof holds for every input that meets the specification’s preconditions and in every execution order. ‘Proved correct’ therefore means the code is correct against this particular specification and its conditions.

MarkTechPost gives an example of how this differs from testing: TensorLean found a data race in FlashAttention-4 that did not show up in 17,760 test runs. Many of the bugs listed on the GitHub page show up only at extreme values, for example as a 32-bit index overflow when the number of elements or offsets approaches or exceeds 2³¹.

Of the 149 kernel instances, or compiled kernels, that were checked, 102 came from FlashAttention. The sources give different bug counts. PANews, summarising CertiK’s press release, and MarkTechPost report 38 bugs. The table on the GitHub page lists 39, because it records three bugs for DeepGEMM. The table below shows the bugs by library.

Library GitHub page MarkTechPost
FlashAttention-2, -3 and -4 23 23
Liger Kernel 7 7
TileOPs 3 3
DeepGEMM (DeepSeek) 3 2
TileLang 2 2
Triton 1 1
Total 39 38

Sources: TensorLean’s GitHub page and MarkTechPost, 21 November 2026. The authors say they have reported all the bugs to the developers.

The counts do not allow a comparison of library quality, because the verification concentrated on FlashAttention.

Volta, from Stanford and Microsoft Research, whose paper appeared on arXiv in November 2025, also checks PTX code. However, it compares a kernel with a reference implementation at fixed tensor sizes and leaves asynchronous operations, such as TMA memory transfers, to future work. TensorLean treats tensor sizes as symbolic parameters and supports the asynchronous operations in FlashAttention-3 and -4, MarkTechPost reported.

For now the results are the authors’ own and cannot be checked independently. The GitHub page says the authors will soon release the paper and open-source the tool together with its PTX semantics, specifications and proofs. MarkTechPost said verification on a 16-core workstation takes a median of about ten minutes and at most 50 minutes. Users of the libraries can find the bug reports, and proposed fixes for some of the bugs, through the links in the GitHub page’s table.

Report an error in this story