LARC
A latent-space implicit reasoning framework for unified multimodal models that performs reasoning in high-dimensional hidden states rather than in discrete token space, mitigating modality mismatch and improving image generation quality and test-time scaling capability.
Timeline (last 90 days)
LARC (Latent Reasoning in Continuous space) is a continuous-space latent reasoning framework that mitigates the token-level bottleneck by performing reasoning over high-dimensional hidden states.
The LARC training pipeline is divided into two stages: curriculum supervised fine-tuning (curriculum SFT) and information-gain reinforcement learning (information-gain RL).
Curriculum SFT provides a smooth transition from discrete to continuous by gradually converting token-level reasoning into latent reasoning, avoiding training instability caused by the model jumping directly into latent space.
Information-gain RL uses the information increment brought by the reasoning process as the reward signal, enabling the model to self-improve on image generation tasks that lack paired supervision.
On multiple text-to-image benchmarks, LARC consistently improves generation quality, prompt following, and compositional fidelity.
On the GenEval benchmark, LARC achieves a 7-percentage-point improvement over BAGEL, reaching state-of-the-art performance.
LARC outperforms token-based generation methods under both parallel and sequential test-time scaling strategies, exhibiting clear test-time scaling behavior.
LARC's improvement on GenEval indicates that the gains from continuous latent reasoning are mainly reflected in dimensions requiring precise semantic alignment such as compositional fidelity, rather than merely overall image aesthetics.
LARC is still at the research stage, and its ideas have general reference value for reasoning enhancement of unified multimodal models.
Future work may explore more complex latent reasoning structures and extend LARC to more continuous modality generation tasks such as video and 3D.
All Facts (10)
LARC (Latent Reasoning in Continuous space) is a continuous-space latent reasoning framework that mitigates the token-level bottleneck by performing reasoning over high-dimensional hidden states.
50%UnverifiedThe LARC training pipeline is divided into two stages: curriculum supervised fine-tuning (curriculum SFT) and information-gain reinforcement learning (information-gain RL).
50%UnverifiedCurriculum SFT provides a smooth transition from discrete to continuous by gradually converting token-level reasoning into latent reasoning, avoiding training instability caused by the model jumping directly into latent space.
50%UnverifiedInformation-gain RL uses the information increment brought by the reasoning process as the reward signal, enabling the model to self-improve on image generation tasks that lack paired supervision.
50%UnverifiedOn multiple text-to-image benchmarks, LARC consistently improves generation quality, prompt following, and compositional fidelity.
50%UnverifiedOn the GenEval benchmark, LARC achieves a 7-percentage-point improvement over BAGEL, reaching state-of-the-art performance.
50%UnverifiedLARC outperforms token-based generation methods under both parallel and sequential test-time scaling strategies, exhibiting clear test-time scaling behavior.
50%UnverifiedLARC's improvement on GenEval indicates that the gains from continuous latent reasoning are mainly reflected in dimensions requiring precise semantic alignment such as compositional fidelity, rather than merely overall image aesthetics.
50%UnverifiedLARC is still at the research stage, and its ideas have general reference value for reasoning enhancement of unified multimodal models.
50%UnverifiedFuture work may explore more complex latent reasoning structures and extend LARC to more continuous modality generation tasks such as video and 3D.
50%