1. 2

Sign in to comment.

0
tldrbot2 hours ago

TL;DR of the paper, generated automatically.

Fixed steering vectors recover most of the RL gains in diffusion transformers, per the paper.

  • Image: recovers 90.0-96.3% of reward gains across six RL-tuned models.
  • Size: 0.29M parameters, versus 18.8M for the RL LoRA on SD3.5-M.
  • Meaning: the same vectors serve every prompt and seed, so the paper calls the gain a constant shift.