Latent Reasoning in Continuous Space for Unified Multimodal Models
arXiv cs.CL · 31 Oct 2026, 03:59 AEST
Unified Multimodal Models struggle with language-based reasoning for image generation due to a modality mismatch. The Latent Reasoning framework in Continuous space (LARC) addresses this issue, improving generation quality and achieving state-of-the-art performance on benchmarks.
Key facts
- Unified 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.
- We attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space.
- To bridge this gap, we introduce a Latent Reasoning framework in Continuous space (LARC), which alleviates the token-level bottleneck by reasoning over continuous hidden states.
- Experiments across diverse text-to-image generation benchmarks show that LARC consistently improves generation quality, prompt following, and compositional fidelity.
- Notably, LARC achieves state-of-the-art performance on GenEval with a 7%p gain over BAGEL and demonstrates clear test-time scaling behavior.
Read aloud
Unified Multimodal Models struggle with language-based reasoning for image generation due to a modality mismatch. The Latent Reasoning framework in Continuous space LARC addresses this issue, improving generation quality and achieving state-of-the-art performance on benchmarks. Unified 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. We attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space. To bridge this gap, we introduce a Latent Reasoning framework in Continuous space LARC, which alleviates the token-level bottleneck by reasoning over continuous hidden states. Experiments across diverse text-to-image generation benchmarks show that LARC consistently improves generation quality, prompt following, and compositional fidelity. Notably, LARC achieves state-of-the-art performance on GenEval with a seven percentp gain over BAGEL and demonstrates clear test-time scaling behavior.
Sources in this cluster
- Latent Reasoning in Continuous Space for Unified Multimodal Models
arXiv cs.CL · 31 Oct 2026, 03:59 AEST