In 2026, still no adversarially robust model on CIFAR-10 (73.7% cap)
maksym_andr · x · 2026-11-22
Maksym Andrusiv notes that in 2026 there's still no adversarially robust model on CIFAR-10. RobustBench's ℓ∞=8/255 leaderboard caps at 73.7% robust accuracy (ICML 2024 scaling-law study using 300M extra synthetic images), leaving a 20-point gap to standard accuracy. RobustBench standardizes evaluation with AutoAttack across 3,000+ papers and open-sources the models.
More from Research
- Open-source Laya: single-pass System 1 decision engine that keeps learning from LLMs — solyarisoftware · 2026-11-22
- Beyond motion tracking: why robots still need closed-loop S0 controllers — animesh_garg · 2026-11-22
- Gary Marcus: without a Lean for biology, AI can't just 'solve biology' — GaryMarcus · 2026-11-22
- Can human minds exit their sensory distribution? Rethinking the baseline for AI creativity — weballergy · 2026-11-22
- Claude did three jobs in the Riemann Hypothesis proof: finding, checking, explaining — Prestigious_Bag_5349 · 2026-11-22
- Kaggle competition pitched to compress LEAN theorem proofs like code golf — MeganRisdal · 2026-11-22