A reusable framework generates verification instances with provably known robustness labels, revealing numeric tolerance issues and bugs in five verifiers while introducing difficulty profiles to diagnose failure modes.
Xiao, and Russ Tedrake
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Stress-Testing Neural Network Verifiers with Provably Robust Instances
A reusable framework generates verification instances with provably known robustness labels, revealing numeric tolerance issues and bugs in five verifiers while introducing difficulty profiles to diagnose failure modes.