Empirical study: logit-level energy guidance on LLaDA-8B masked diffusion (13 experiments, +39% semantic steering). Extends Diffusion-LM classifier guidance to 8B scale, gradient-free.
natural-language-generation adversarial ai-safety diffusion-models energy-based-models controllable-generation llm-security diffusion-lm llm-steering masked-diffusion llada logit-guidance
-
Updated
Jul 4, 2026 - Python