Encoder-Decoder
Also called Seq2Seq
A two-part design where one stack reads the input and another writes the output, linked by attention.
Think of it like
A translator with a listening phase and a speaking phase — understand fully first, then produce.
Example
T5 and classic translation models use this: the encoder digests the source, the decoder emits the translation.
How it actually works
The original transformer was encoder-decoder: the encoder builds a representation of the input, and the decoder generates the output while cross-attending to that representation. It fits tasks with a clear input-to-output mapping like translation or summarization. Many modern chat models skip the encoder and go decoder-only, but the two-part design still shines for structured transduction.
For product teams
The classic fit for “turn this input into that output” tasks like translation and summarization.
For engineers
An architecture pairing a bidirectional encoder with an autoregressive decoder joined by cross-attention.
Related
- Encoder — Combines an encoder and decoder.
- Cross-Attention — Linked via cross-attention.
- Transformer — The original transformer shape.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome