aiqbench analysis
This arXiv preprint proposes LARC, a training method that lets unified multimodal models reason over continuous hidden states instead of discrete tokens before generating images. Its reported gains rest on benchmark results, including a 7-point GenEval improvement over BAGEL, and the claimed test-time scaling behavior is the main thing to verify against the full paper. Readers tracking image generation should note that these are self-reported results on specific benchmarks, not evidence of broad real-world improvement.
Latent Reasoning in Continuous Space for Unified Multimodal Models
arXiv:2610.42149v1 Announce Type: cross Abstract: 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. As a result, existing reasoning-augmented UMMs often fail to inherit the test-time scaling and fine-grained controllability observed in large language models. 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. Specifically, the training framework consists of two stages: (i) curriculum supervised fine-tuning (SFT), and (ii) information-gain reinforcement learning (RL). First, the curriculum SFT stage gradually converts token-level reasoning into latent reasoning. Subsequently, the information-gain RL stage self-evolves latent reasoning without ground truth supervision. 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, outperforming token-based generation under both parallel and sequential test-time scaling strategies.
Abstract: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. As a result, existing reasoning-augmented UMMs often fail to inherit the test-time scaling and fine-grained controllability observed in large language models. 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. Specifically, the training framework consists of two stages: (i) curriculum supervised fine-tuning (SFT), and (ii) information-gain reinforcement learning (RL). First, the curriculum SFT stage gradually converts token-level reasoning into latent reasoning. Subsequently, the information-gain RL stage self-evolves latent reasoning without ground truth supervision. 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, outperforming token-based generation under both parallel and sequential test-time scaling strategies.
Published as conference proceeding for NeurIPS 2026. Last two authors advised equally
Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2610.42149 [cs.CV]
(or arXiv:2610.42149v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2610.42149 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 305 words
Related coverage
Regularizing modality contribution drift in multimodal continual learning
arXiv · Jul 31, 2026
When Modalities Remember: Continual Learning for Multimodal Knowledge Graphs
arXiv · Aug 6, 2026
SiGMA: Sign-Guided Merging and Adaptation for Multimodal Continual Instruction Tuning
arXiv · Jul 24, 2026
CoRe-MoE: Compact Reusable MoE for Continual Multimodal Instruction Tuning
arXiv · Sep 1, 2026