Rethinking in Spikes: Mitigating Hallucinations in MDLMs with Step-Aware Decoding

Media reports Hugging Face Papers Source language: Chinese

论文摘要(作者报告,未经本站复测):Recent advancements in multimodal diffusion language models (MDLMs) have exhibited strong global modeling capabilities through iterative refinement.

Full text

StepRefine, a plug-and-play decoding strategy, mitigates hallucinations in multimodal diffusion language models by switching between discrete denoising and latent refinement based on entropy spikes and distributional stability.

Sources

← Back to Daily AI Report