Encoder-decoder

Encoder-decoder is the architecture from the original 2017 "Attention Is All You Need" paper. The encoder is a stack of Transformer blocks that processes the input bidirectionally — every token can see every other token. The decoder is another stack that generates the output one token at a time, with two kinds of attention: causal self-attention over what it has generated so far, and cross-attention that reaches into the encoder's output to pull in source information. This architecture dominated machine translation and structured sequence-to-sequence tasks. T5, BART, mT5, and the original Transformer for English-to-German translation are all encoder-decoder. It's a clean fit when there's a real input-output split: the input gets encoded once into a dense representation, and the decoder repeatedly attends to that representation while generating. For translation, that's exactly the shape of the problem. So why aren't chat models encoder-decoder? A few reasons. Decoder-only models scale more cleanly on raw text (no need for paired input-output examples). They handle multi-turn chat naturally — the whole conversation is just one growing sequence. And once decoder-only models got big enough, they matched encoder-decoder performance on translation and summarization too, while being more flexible for everything else. By 2022 the field had largely moved on. Encoder-decoder still shows up in production for narrow workloads. Google Translate uses a custom encoder-decoder. Many specialized summarization, OCR (like TrOCR), and speech-to-text systems (Whisper) are encoder-decoder. But for general-purpose AI assistants, decoder-only won, and the term "LLM" almost always means decoder-only now.