Researchers from KAIST AI propose LARC, which lets unified multimodal models reason over continuous hidden states instead of discrete tokens before generating an image.
Training: Two stages. Curriculum supervised fine-tuning gradually turns token-level reasoning into latent reasoning, then information-gain reinforcement learning self-evolves the latent reasoning without ground truth supervision.
Results: The paper says LARC improves generation quality, prompt following and compositional fidelity across text-to-image benchmarks, and reaches state-of-the-art on GenEval with a 7%p gain over BAGEL.
Scaling: According to the abstract, it shows clear test-time scaling and beats token-based generation under both parallel and sequential strategies.
Sign in to comment.
TL;DR of the paper, generated automatically.
Researchers from KAIST AI propose LARC, which lets unified multimodal models reason over continuous hidden states instead of discrete tokens before generating an image.