Paper Reports
Plain-language walkthroughs of papers from our group — what problem the paper attacks, how the system works, and what the numbers actually say. Each report links back to the paper itself.
2026 3 reports

SC '26
ZOCheck: CPU-Shadow Checkpointing for Zeroth-Order LLM Fine-Tuning
A zeroth-order training step is a seed and a scalar. Log those, let a CPU shadow replay them, and checkpointing costs almost nothing while recovery is bit-for-bit exact.
219.7× lower checkpoint overhead than asynchronous full-state checkpointing21.3× less end-to-end wasted time on Qwen3-8B across MTBFs of 3–15 h1.55× faster recovery after a software failure than async checkpointing every 7 steps; 586× vs. log-only

SC '26
Understanding Reliability Bottlenecks in LLM Agents for HPC I/O Telemetry Analysis
We gave LLM agents 49 real I/O questions about the Polaris supercomputer, three tries each. What makes them reliable depends on the question: curated context for extraction, a reviewer for checkable analysis, and still an expert for the hard, open-ended ones.
49 expert-curated tasks over 2+ years of Polaris telemetry73.0 → 93.5% extraction score with curated schema context+13.2 pts from a process-aware reviewer on analysis tasks

Cluster '26
LayerCheck: Adaptive Layer-wise Checkpointing for Large Language Model Post-training
Save only the layers that changed, and rebuild the rest at recovery — 22.6× smaller checkpoints without changing what the model learns.
22.6× smaller total checkpoint storage than LowDiff0.54% max loss deviation after recovery, within training noise3.4× faster recovery at matched restart freshness
2025 1 report

SC '25
STELLAR: Storage Tuning Engine Leveraging LLM Autonomous Reasoning for HPC Parallel File Systems
An agentic LLM tuner that reaches near-expert Lustre configurations in five runs instead of thousands.
7.8× peak speedup over default Lustre settings≤ 5 configurations tried per tuning run13 parameters kept, out of 159+ Lustre exposes