LARC: latent reasoning in continuous space tops GenEval with +7p over BAGEL

burny_tech · x · 2026-11-10

Researchers from KAIST, POSTECH and RLWRLD introduce LARC (Latent Reasoning in Continuous Space), targeting a core bottleneck in Unified Multimodal Models (UMMs).

Diagnosis: reasoning happens in discrete token space while image generation lives in continuous space—a fundamental modality mismatch that stops language-style chain-of-thought from improving generation.
Method: reason directly over continuous hidden states, trained in two stages—curriculum SFT gradually converts token-level reasoning into latent reasoning, followed by information-gain RL that self-evolves without ground-truth supervision.
Results: state-of-the-art on GenEval with a 7-percentage-point gain over BAGEL, plus consistent improvements in prompt following and compositional fidelity, and clear test-time scaling behavior beating token-based generation under both parallel and sequential strategies.

Original post →

LARC: latent reasoning in continuous space tops GenEval with +7p over BAGEL

More from Multimodal