Authors: Maksim Melnichenko (University of California, Berkeley; International Computer Science Institute), Jack Dongarra (University of Tennessee, Knoxville; Oak Ridge National Laboratory (ORNL)), James Demmel (University of California, Berkeley), Piotr Luszczek (University of Tennessee, Knoxville), Hartwig Anzt (Technical University of Munich; University of Tennessee, Knoxville), Michael Mahoney (University of California, Berkeley; International Computer Science Institute), Rasmus Larsen (NVIDIA Corporation), Riley Murray (Sandia National Laboratories), Vasileios Georgiou (Purdue University), Kimon Fountoulakis (University of Waterloo), Petros Drineas (Purdue University), Junchao Zhang (Argonne National Laboratory), Tze Meng Low (Carnegie Mellon University)
Abstract: Since SC25, a revolution has taken hold: coding agents have moved from curiosity into nearly every institution's pipeline, and "vibe coding" is now how much software, numerical included, gets written. That matters for numerical linear algebra (NLA), whose libraries underpin modern science and whose trustworthiness was built over 50 years of human expert engineering. Now, when a plausible NLA kernel is nearly free to generate, what holds that trust together? This BoF gathers established leaders and younger NLA scientists from academia and industry to open a discussion, so our community shapes this shift rather than being swept along.
Long Description: The world of scientific research has entered the age of AI-assisted development. Adoption is nearly universal: almost every laboratory and university now provides staff with commercial coding agents, and DOE's flagship Genesis Mission points the national laboratories squarely at AI for science. Numerical linear algebra (NLA) is no exception. Through various projects, the agents have reached our kernels directly, generating numerical routines that increasingly rival established libraries while handling the scaffolding, tests, and ports that once cost weeks. For NLA, this is both an opportunity and a hazard, and this BoF aims to discuss both aspects.
Algorithms in this field are published and still actively invented: batched and mixed-precision kernels, randomized methods, and whatever an agent may propose tomorrow. But knowing an algorithm has never been the scarce resource; trusting an implementation of it is, and that trust comes from the test and verification suite that shows the implementation is correct and stable. Each such suite lives inside a single library, entangled with its code; the field has no shared, library-independent form of that certifying layer, and tests, once fixed, can be gamed by a clever implementation, human or machine. When plausible kernels can come from anywhere at almost no cost, what should certification look like? That is a question we want this BoF to open, not one we claim to have answered.
Coding agents did not create this gap, but they've widened it. Agents amplify existing practices rather than replace judgment: they make good habits faster, and bad habits faster too. What they amplify in NLA is ours to choose, and a handful of concrete questions decide it. What's the future role of our hard-won standards and interfaces? What tooling would let us certify what agents produce? Who maintains agent-generated code, if anyone, and who gets credit? What will expertise mean for the next generation of numerical software developers? Where is the field's next frontier?
Our BoF session is built to work through these together. We'll discuss the questions in groups working in parallel, with two leads for each group: an established leader who has helped build this field, and a younger researcher as co-facilitator, drawn from academia or industry and working in an adjacent area such as formal verification, optimization, GPU and mixed-precision computing, or research software engineering. In the plenary part of the session, each lead will introduce the discussion topic with a short, three-minute prompt. The pairs will then invite session attendees to participate in parallel small group discussions. We'll close by bringing the groups back together to hear each co-facilitator report on their group's discussion: where agents belong in NLA, and which open problems are worth pursuing next.
We invite everyone who builds, maintains, funds, or depends on numerical software, from numerical analysts and library developers to research software engineers and representatives of AI companies, whether you arrive enthusiastic, skeptical, or simply curious about where our field is headed.