LARC enables continuous latent reasoning for multimodal image generation
Original title: Latent Reasoning in Continuous Space for Unified Multimodal Models
LARC uses curriculum SFT and information-gain RL to bridge discrete token and continuous image spaces, delivering 7% GenEval gains and test-time scaling.
Unified Multimodal Models struggle to leverage language chain-of-thought for image generation due to modality mismatch. LARC reasons over continuous hidden states in two stages: curriculum supervised fine-tuning followed by information-gain reinforcement learning. Experiments show improved quality, prompt following, and compositional fidelity across text-to-image benchmarks.
- Curriculum SFT converts token-level reasoning to latent reasoning
- Information-gain RL self-evolves latent reasoning without ground truth
- 7%p GenEval gain over BAGEL
- Outperforms token-based generation under parallel and sequential test-time scaling
From the original article
1 min read at arxiv.orgUnified Multimodal Models (UMMs) achieve remarkable success on diverse multimodal understanding and generation tasks, yet they still struggle to effectively leverage language-based chain-of-thought reasoning for image generation. As a result, existing reasoning-augmented UMMs often fail to inherit the test-time scaling and fine-grained controllability observed in large language models.
By Byungwoo Jeon · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.