BRAID: Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process

The paper introduces BRAID, a reinforcement-learning framework that formulates interleaved text-image-text reasoning as a unified Markov decision process. Instead of optimizing text with RL while treating image generation as a supervised surrogate, BRAID uses a shared trajectory-level advantage for both text tokens and image denoising paths, with modality-native policy gradients. A vision-language-model judge additionally scores intermediate images according to their utility for reasoning, providing denser turn-level feedback for long-horizon credit assignment. The authors report consistent improvements over multiple baselines on spatial reasoning and visual perception benchmarks, although the abstract does not provide numerical results or implementation details.