How does CURE perform across LLaDA-8B-Instruct,Dream-v0-Instruct-7B, and LLaDA-1.5 on 5 major jailbreak benchmarks (Wild Jailbreak, JBB-Behaviors, HarmBench, PAD-AdvBench, DIJA-AdvBench)?
🛡️Crushing Jailbreak Attacks (Lowest ASR Across All Models)：
• Cuts average ASR from 47.20% down to 1.36%(1.83%on LLaDA-8B,2.19%on LLaDA-1.5, and near-zero 0.06% on Dream-7B!)
• Excels against early-anchor attacks (PAD & DIJA), slashing LLaDA-8B ASR from 93.65% / 98.65% down to 4.42% / 4.23%.
📈Pushing the Safety–Utility Pareto Frontier:
• On MMLU & GSM8K, CURE preserves base capabilities without over-refusal—boosting LLaDA-8B defense success from 51.00% to 98.17%while increasing average utility accuracy from65.68% to 68.48%!