Test-Input Generation for Tensor Programs: What Actually Finds Kernel Bugs

Collection of recent arXiv CS/SE papers (June 29, 2026) covering testing, robustness, and evaluation in software and ML systems. Key security-relevant findings: for tensor/GPU kernels, boundary-only shape sampling found 78% of seeded kernel bugs with 0% false positives while adversarial value sampling reached 99% recall but produced widespread false positives by injecting NaN/Inf inputs; this suggests test-input choices materially affect bug discovery and spurious failure signals. Other notable papers survey GenAI pedagogy using Bloom's taxonomy, systematic use of formal grammars in BPM, large