LARC runs chain-of-thought in continuous space for image gen

A new preprint introduces LARC, a Latent Reasoning framework in Continuous space, and reports a 7-percentage-point gain over BAGEL on GenEval. Posted on arXiv, the paper argues that reasoning-augmented unified multimodal models have been failing for one specific reason, and the authors name it directly.

"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," the authors write.

The training runs in two stages. A curriculum supervised fine-tuning stage "gradually converts token-level reasoning into latent reasoning." Then an information-gain reinforcement learning stage "self-evolves latent reasoning without ground truth supervision." The claim: reasoning done entirely in the continuous latent space the image model already inhabits removes the token-level bottleneck that has hobbled prior attempts.

The headline number is that 7-percentage-point jump over BAGEL on GenEval, which the paper calls "state-of-the-art." The authors also report "clear test-time scaling behavior" for image generation under both parallel and sequential strategies, outperforming token-based reasoning at matched compute.

The abstract publishes no per-category GenEval breakdown, no inference-cost comparison, and no results against proprietary generators. It is one preprint from one group.