Latent Refinement Decoding: Enhancing Diffusion-Based Language Models by Refining Belief States
Published in NeurIPS 2026
Latent Refinement Decoding (LRD) combines latent refinement with a predictive feedback loop to preserve information at masked positions and coordinate token commitments. On coding and reasoning benchmarks, it improves accuracy while delivering decoding speedups of up to 10.6×.
Recommended citation: Qinglin Zhu, Yizhen Yao, Runcong Zhao, Yanzheng Xiang, Amrutha Saseendran, Chen Jin, Philip Teare, Bin Liang, Yulan He, Lin Gui. 2026. "Latent Refinement Decoding: Enhancing Diffusion-Based Language Models by Refining Belief States." In NeurIPS 2026.
Download Paper