prefixlm-hybrid-mask-bidirectional-prefix-causal-generation

Summary

PrefixLM lets a single decoder handle both reading and writing: the input text gets full two-way attention so the model can understand the whole context at once, while the output tokens are generated one-at-a-time looking only backward. In practice, this means you get encoder-decoder quality on tasks like text infilling or translation without needing a separate encoder block, keeping the architecture simpler and easier to serve.