Efficient Reinforcement for Visual-Textual Thinking with Discrete Diffusion Model

The paper introduces a novel approach using multimodal discrete diffusion models for reinforcement learning in visual-textual reasoning, offering an alternative to traditional autoregressive models. This method achieves a 26.9% reduction in computation during visual rollouts and demonstrates significant performance improvements through a factorized reward assignment strategy, yielding an 11.2% enhancement over joint reward assignment and a 38.04% gain over baseline models. This advancement is crucial for practitioners as it enhances efficiency and effectiveness in developing unified multimodal systems capable of interleaved reasoning without the computational overhead of full image regeneration.