keyword

Non-autoregressive conditional sequence generation

Non-autoregressive conditional sequence generation is a machine learning framework that produces an entire target sequence simultaneously in parallel based on a conditioning input sequence, rather than generating elements one by one sequentially. Unlike standard autoregressive models that condition the prediction of each subsequent token on previously generated tokens, non-autoregressive models break this strict sequential dependency to dramatically reduce inference latency during generation. Commonly applied in sequence-to-sequence tasks such as machine translation, speech synthesis, and text summarization, this paradigm achieves faster decoding while employing specialized architectures, latent variables, or iterative refinement mechanisms to handle cross-token dependencies and preserve coherence in the generated output.

1 item