About this paper
A fixed set of steering vectors added to a frozen diffusion transformer recovers most of the quality gains from reinforcement learning post-training, indicating that much of the improvement comes from a constant shift rather than prompt-specific transformations.
Reinforcement learning (RL) post-training substantially improves the visual quality, text rendering, and compositional accuracy of diffusion transformers. Because capabilities such as rendering requested text or composing multiple objects are inherently prompt-dependent, a natural hypothesis is that improving them requires learning complex, input-dependent transformations. To isolate how much of these gains an input-independent shift explains, we probe each RL-tuned model with a deliberately simple parameterization: a fixed set of steering vectors added to the block outputs of the frozen base model. On SD3.5-M, these vectors hold 65x fewer parameters than the RL-tuned LoRA (0.29M against…
reinforcement learning post-trainingdiffusion transformerssteering vectorsblock outputsfrozen base modelSD3.5-MRL-tuned LoRAvelocity predictionsRL-tuned checkpointspolicy-gradient RL
- Upvotes
- 8
- Authors
- 6
- Published
- Nov 19, 2026
Summary and abstract excerpt via Hugging Face Papers ↗ · arXiv.
Trending history
RL Post-Training in Diffusion Transformers Is Activation Steering was first seen on Hugging Face · Daily papers on Nov 20, 2026. Since then it has appeared in 23 checks on 3 different days. Its best position was #14 of 30, reached on Nov 22, 03:29 UTC. It is #14 right now, holding steady at #14 over the last 6 checks.