Concept Contrast Fable

A two-module architecture that reads everything before writing, and a single-module architecture that writes while only looking at the preceding context

Conclusion first:The encoder-decoder architecture has two parts: the encoder digests the input sequence into a mature internal representation, and the decoder expands that internal representation into an output sequence. A pure decoder architecture has only one part, and each time it generates the next token it refers only to previously generated tokens and the original prompt. The former builds a global understanding of the input internally before generating, while the latter writes as it reads, with no separate digestion step. Both architectures generate text, but the 'way of understanding the input' is completely different, which determines that translation and summarization are better suited to encoder-decoder, while free conversation is better suited to pure decoder.

The encoder-decoder architecture has two parts: the encoder digests the input sequence into a mature internal representation, and the decoder expands that internal representation into an output sequence. A pure decoder architecture has only one part, and each time it generates the next token it refers only to previously generated tokens and the original prompt. The former builds a global understanding of the input internally before generating, while the latter writes as it reads, with no separate digestion step. Both architectures generate text, but the 'way of understanding the input' is completely different, which determines that translation and summarization are better suited to encoder-decoder, while free conversation is better suited to pure decoder.

Fable

Two translation workflows in the translation house: swallowing the whole text and then spitting it out, or translating sentence by sentence as you listen

The capital's translation bureau had two teams of translators. Team A, after taking on a job, would first silently read the entire text from beginning to end twice, drawing a relationship map in their minds—who had a grudge against whom, which event happened first, which piece of evidence in the earlier text the conclusion of the final paragraph was tied to. Only after the map was fully drawn did they begin translating. During translation, they often moved content from later sections to the beginning, because the original narrative order differed from the conventions of the target language, but the meaning could not be distorted.

Class B's style was completely different. They started from the beginning, translating the first sentence as soon as they read it, then moving to the second. They never looked at later text—after all, it hadn't been translated yet, and there was no need to. For the sentence "He walked toward the city called London," Class B finished translating as soon as they read "city." After finishing, they continued to the next sentence without going back to adjust earlier wording.

Team A and Team B performed similarly when translating letters—a three-hundred-word letter with a greeting at the start, the main content in the middle, and a signature at the end naturally lines up in order. But when they encountered a meeting minutes document, the difference exploded. The minutes were written in reverse order: first the voting result from the third day, then back to the first day for background. After reading all three days of content, Team A straightened out the timeline when translating—the first day placed first. Team B translated in the original order, so the third day's vote came out first, leaving readers baffled: how did this person vote before the meeting even started?

Another long official document requiring a concise summary arrived. Team A read the full text first and grasped the core—it was a funding request, and the key reason was "the granary is leaking." So they wrote the summary around the leak. Team B translated it sentence by sentence into a summary from the beginning; the first two paragraphs were about the yamen's organizational structure, and Team B treated these as the main content. By the time they reached the leak paragraph, there was no space left. The summary became an institutional introduction rather than the reason for the funding request.

But there were also times when Team A lost out. A live conversation transcription—two people were arguing, and every sentence rebutted the previous one. Team B translated sentence by sentence as it listened, responding quickly and without lag, and the flow of conversation felt natural. Team A insisted on waiting until the other side finished a full turn and digested it before translating, arriving three seconds late. The rhythm of the argument was disrupted, and the speaker thought the interpreter was hesitating.

The translation house later established a new rule: letters and dialogues go to Team B, because order itself is part of the content and does not need to be reorganized. Official documents, minutes and papers go to Team A, because the input needs to be understood as a whole before output, and the logical structure of the entire piece matters more than the fluency of individual sentences.

On the title page of the regulations, the old dispatcher wrote a sentence: The difference is not in whether you translate well, but in whether you must make an overall judgment about the full text before translating. Class A said that writing without making a judgment is gambling; Class B said that by the time you make your judgment, the customer across the way has already left—each has its own kind of frankness.

Encoder-decoder

An architecture composed of two modules, an encoder and a decoder. The encoder processes the entire input sequence into an internal representation, and the decoder then generates the output sequence step by step based on this global representation. Typical representatives are T5 and BART, which excel at tasks requiring global understanding such as translation and summarization.

Decoder-only

An architecture containing only a decoder module, using self-attention to focus on generated tokens and the input prompt. Each time it predicts the next token, it relies only on the preceding text. The representative is the GPT series, which excels at free-form text generation and dialogue.

Story Mapping

Class A reads the full text before writing: encoder-decoder, understand globally first, then generate output
Class B reads one sentence and translates one sentence: pure decoder, autoregressively generating token by token
Two translations of meeting minutes in reverse order: tasks that require reorganizing information expose the limitations of Class B
Class B document summarizers catch the wrong key points: global understanding is essential when input has cross-paragraph dependencies
Real-time dialogue Class A was three seconds slow: pure decoders are more suitable when low-latency streaming responses are needed
Translation bureau new rule: partitioned tasks—two architectures each have strengths; task type determines architecture choice

Put it into model selection and task matching

Tasks that require deep understanding of the input and reorganization—translation, long-document summarization, structured information extraction—benefit from the global encoding capability of encoder-decoder architectures, which yields more coherent results. Tasks that require fluent generation and fast response—dialogue, creative writing, code completion—are more natural with the autoregressive generation of pure decoders. In reality, most dialogue scenarios have already been taken over by pure decoders, but when you encounter business scenarios that require "thinking it through before writing," do not forget that encoder-decoder models such as T5 and BART still lead on certain metrics. Before choosing a model, first judge whether the input needs to be digested as a whole.

Encoder-decoder Decoder-only