Latent Reasoning in Continuous Space for Unified Multimodal Models

TL;DR

The paper proposes LARC, a framework that enables continuous-space reasoning in unified multimodal models by shifting from discrete token-based reasoning to latent reasoning over continuous hidden states. It uses a two-stage training process involving curriculum supervised fine-tuning and information-gain reinforcement learning to improve image generation quality, prompt adherence, and compositional accuracy.

Why It Matters

By aligning reasoning with the continuous latent space used in image generation, LARC enables test-time scaling and fine-grained control similar to large language models, overcoming a key limitation in current multimodal systems. This advances the development of more capable and controllable generative AI for complex visual tasks.

Key Findings

• LARC improves text-to-image generation quality and prompt following across multiple benchmarks.
• The model exhibits clear test-time scaling under both parallel and sequential strategies, unlike token-based approaches.
• Information-gain reinforcement learning enables self-evolution of latent reasoning without relying on ground truth annotations.