Course contentsShow
← Machine Learning and Deep Learning
Lesson 1102 of 3,53824. The Transformer ArchitectureFree lesson

Encoder-Decoder vs Decoder-Only Trade-offs

When to use encoder-decoder: translation and tasks requiring separate encoding of input context.

Encoder-Decoder vs Decoder-Only Trade-offs

What you'll learn: When to choose encoder-decoder architectures versus decoder-only models based on your task requirements.

The Core Decision

Now that you understand both encoder-decoder and decoder-only architectures, you need to know when to use each. The key distinction lies in whether your task requires separate encoding of input context.

When to Use Encoder-Decoder

Translation and structured input-output tasks are ideal for encoder-decoder models:

  • Machine Translation: You have a complete source sentence in one language that needs to be fully understood before generating the target language. The encoder processes the entire French sentence, then the decoder generates English.
  • Summarization: The full document must be encoded before creating a condensed version
  • Question Answering: The context/passage is encoded separately from generating the answer

Why it works: The encoder can use bidirectional attention to understand the entire input without worrying about autoregressive constraints. The decoder then focuses purely on generation while attending to this rich encoded representation via cross-attention.

When to Use Decoder-Only

Open-ended text generation often works better with decoder-only:

  • Text completion: No clear separation between "input" and "output"
  • Conversational AI: Context and response blend together
  • Simpler training: One unified architecture, no cross-attention complexity

The trade-off: Decoder-only models process everything autoregressively (left-to-right with masking), which can be less efficient for encoding fixed inputs, but scales better and is architecturally simpler.

Key Takeaway: Use encoder-decoder when you have distinct input and output with full input context available (translation, summarization), and decoder-only for unified text generation tasks where input/output boundaries blur.