ASK 2027
Workshop on Agentic Methods for Sparse and Irregular HPC Kernels

Overview

ASK is a half-day workshop on agentic large-language-model (LLM) workflows for high-performance computing kernels, centered on sparse and irregular kernels. Its premise is the distinction between generation and agency. A generative system emits a candidate kernel from a prompt. An agentic system selects a data structure, compiles, executes on real hardware, reads performance counters, diagnoses the bottleneck, and revises. Early systems now demonstrate the promise of this closed loop, but methods and evaluation remain fragmented—especially for sparse and irregular workloads.

Sparse kernels make this question concrete. The decisions that govern their performance are largely invisible in the source: which representation to store the data in, how to reorder and partition it, and how much preprocessing to invest against the reuse it buys. These decisions depend on the input as much as on the code, interact with one another, and rarely admit a single best answer. That is an unusually good match for an agent that can try something, measure the result, and revise, and we expect it to produce workflow designs quite different from those developed for the dense case.

The workshop brings together the sparse and combinatorial kernel community and the agentic LLM community, which today publish in largely disjoint venues. We welcome workflow and system design, empirical studies of existing agents, benchmarks and evaluation methodology, compiler and tooling support, and experience reports. Negative results and failure analyses are explicitly in scope.

Call for Papers

ASK 2027 invites original papers on topics including:

• Agentic loops for kernel optimization: compile, run, profile, revise; counter- and roofline-guided search; autonomous autotuning
• Sparsity format and algorithm selection by LLM agents; input-dependent specialization
• Multi-agent decomposition of HPC tasks (generate, verify, tune, port)
• Tool use and grounding: compilers, profilers, schedulers, performance models, and sparse DSLs (e.g., TACO, MLIR SparseTensor) as agent tools
• Correctness and numerical verification of agent-written kernels
• Agentic performance portability across CPUs, GPUs, and accelerators
• Benchmarks, datasets, and evaluation methodology for HPC coding agents
• Cost, energy, and human-in-the-loop trade-offs
• Negative results and failure analyses

Dense, stencil, and FFT kernels are in scope when the contribution is the agentic method rather than the kernel itself.

Full papers may be up to 10 pages in IEEE conference format. A short-paper track of up to 4 pages is open to work in progress; short papers are presented at the IPDPS cross-workshop poster session. Submissions are made through Linklings and reviewed single-blind, with at least three reviews per paper. All accepted papers are peer-reviewed and appear in the IPDPS workshop proceedings (IPDPSW) for inclusion in the IEEE Xplore Digital Library. The workshop is held in person, and at least one author of each accepted paper must register and present in person.

Because agentic results vary between runs, submissions should report model version, sampling parameters, and the number of trials behind any reported number. Submission of artifacts is encouraged. The submission link and final deadline information will be posted shortly.

Important Dates

Paper deadlines are planned and remain subject to final confirmation with IPDPS. Submission deadlines are Anywhere on Earth. The June 2 workshop date is confirmed.

Program

ASK 2027 will be a four-hour morning workshop featuring an invited talk, peer-reviewed technical presentations, and a moderated discussion on evaluation methodology for agentic HPC workflows. The invited speaker and detailed schedule will be announced after confirmation.

The program committee is currently being formed and will be published here once invitations are complete.

Organizers

• Ramakrishnan Kannan, Oak Ridge National Laboratory, USA
• Yongseok Soh, Oak Ridge National Laboratory, USA
• Ariful Azad, Texas A&M University, USA