A developer reported achieving a 232-fold speedup over a baseline GPU kernel after using coding agents in an automated research loop during a 14-day optimization contest. In an account highlighted on August 15, 2026, the participant said the result placed 12th among 183 entrants in a GPU Mode competition organized with Core Automation.
The challenge involved batched square compact-Householder QR factorization for FP32 CUDA matrices. Submissions had to return the same compact representation as PyTorch’s `torch.geqrf`: a matrix whose upper triangle contains the triangular factor R, whose lower portion stores Householder vectors, and a vector of reflector coefficients called tau. The checker rebuilt Q, extracted R and tested whether the factors accurately reconstructed the input.
Correct entries were ranked by geometric-mean runtime across matrix shapes and conditioning cases. Important dimensions included batches of 512-by-512 matrices and larger 1,024, 2,048 and 4,096 cases. Lower-precision formats could be used internally, but the returned factors still had to satisfy checks designed for FP32 QR results. This prevented speed improvements from simply discarding the required numerical accuracy.
The contest’s command-line tooling let agents test, benchmark and submit directly, returning performance information for individual shapes as well as an overall score. That produced the tight feedback loop behind the author’s experiment. Across two weeks, the participant made more than 1,500 submissions, though rapid collective use sometimes lengthened queues and exhausted hosted-compute credits.
Before the contest, the author had about a year of familiarity with GPU kernel optimization but had not worked in the field professionally. Early work included learning QR decomposition and Householder reflections with AI assistance and educational videos. The eventual architecture used blocked Householder operations and a trailing WY update, an approach intended to turn a sequential factorization problem into work better suited to matrix multiplication on GPUs.
The account calls the process auto-research, while acknowledging that others may call it loop engineering. Its central claim is not that an agent independently discovered the whole solution. Domain knowledge helped the participant ask better questions, recognize useful terminology and evaluate proposed approaches. At the same time, deterministic correctness tests and benchmark feedback gave the agent concrete signals for revising code.
The reported 232-fold improvement is relative to the contest baseline, not a universal comparison with production QR libraries. It comes from the participant’s own result and leaderboard account, with no independent reproduction in the supplied evidence. Even so, the experiment illustrates a specific pattern for agent-assisted optimization: constrain the output precisely, automate evaluation, return granular feedback and iterate many times. The ranking also shows that large gains over a baseline did not necessarily produce a top-ten result in a competitive field.
