Latent Action Control for Reasoning-Guided Unified Image Generation

In three linesLAC (Latent Action Control) makes reasoning actionable in unified generative models by representing planning and diagnosis as continuous hidden actions. Integrated into BAGEL-7B-MoT, LAC improves compositional and knowledge-grounded generation via variational alignment and GRPO, with major gains on spatial relations and attribute binding.

Summary generated by Claude — human-verified