Meet CURE: Aligning Diffusion Language Models via Counterfactual Token Re-masking—accepted as a #NeurIPS2026 Oral (Top 0.36%, 112/30,709)🎉
Can seemingly harmless words like "might", "by", or "to" jailbreak a Diffusion LLM—and can we surgically reverse the trajectory at test time with just 0.7% extra compute?
Without tuning base model parameters, CURE brings test-time alignment to Diffusion LLMs (LLaDA-8B, LLaDA-1.5, Dream-7B)—slashing average jailbreak Attack Success Rate (ASR) from 47.20% to 1.36% while sitting squarely on the Safety–Utility Pareto frontier 👇