Paper: RL post-training gains in diffusion transformers are mostly activation steering

Tuna Han Salih Meral · hf · 2026-11-21

A paper from Virginia Tech, fal, Georgia Tech and IBM Research shows most quality gains from RL post-training of diffusion transformers come from a constant shift: a fixed set of steering vectors (0.29M params, 65× fewer than RL-tuned LoRA) recovers 90.0–96.3% of reward gains across six RL-tuned image models and 89.8%/111.9% on two video models, and beats weight-based RL on OCR text rendering and PickScore on SD3.5-M. Policy-gradient RL largely steers the pretrained generator toward what it can already compute.

Original post →

More from Multimodal

Multimodal channel →