Tencent WeChat AI Team Launches New Diffusion Language Model WeDLM to Improve Reasoning Efficiency

Tencent WeChat AI team has launched a new diffusion language model called WeDLM, aimed at improving text generation efficiency. The model combines diffusion models with causal attention mechanisms, and uses topological reordering technology to be compatible with KV caching, solving the issue of inference efficiency caused by bidirectional attention in traditional diffusion models, and breaking through the limitations of large models such as GPT in parallel inference.