Activation Steering for Masked Diffusion Language Models
Published:Dec 30, 2025 11:10
•1 min read
•ArXiv
Analysis
This paper introduces a novel method for controlling and steering the output of Masked Diffusion Language Models (MDLMs) at inference time. The key innovation is the use of activation steering vectors computed from a single forward pass, making it efficient. This addresses a gap in the current understanding of MDLMs, which have shown promise but lack effective control mechanisms. The research focuses on attribute modulation and provides experimental validation on LLaDA-8B-Instruct, demonstrating the practical applicability of the proposed framework.
Key Takeaways
Reference
“The paper presents an activation-steering framework for MDLMs that computes layer-wise steering vectors from a single forward pass using contrastive examples, without simulating the denoising trajectory.”