decoder-only-won-the-scaling-race

IN derived (depth 1)

Created 2026-06-21T09:52:14+00:00 · Reviewed 2026-06-21T14:41:08+00:00

Decoder-only architecture became the dominant paradigm for large language models, with GPT, Llama, and most frontier models choosing autoregressive causal modeling over encoder-only or encoder-decoder variants.

Summary

The industry converged on the "predict the next token one at a time" design (used by GPT, Llama, and most frontier models) over the bidirectional or read-then-generate alternatives, meaning virtually every major language model today shares the same fundamental training logic. This simplifies the ecosystem and tooling, but it also locks in specific tradeoffs around how models use context and what kinds of tasks they handle most naturally.

Justifications

SL — Of three transformer variants, decoder-only prevailed for generative scaling while encoder-only (BERT) remained specialized for understanding tasks

Antecedents (all must be IN):

  • IN llama-decoder-only-autoregressive — Llama is a decoder-only autoregressive transformer architecture (like GPT-3), not encoder-decoder or encoder-only
  • IN llama-decoder-only-autoregressive — Llama is a decoder-only autoregressive transformer architecture (like GPT-3), not encoder-decoder or encoder-only
  • IN transformer-three-architecture-variants — The three main Transformer architecture variants are encoder-only (e.g., BERT, bidirectional), decoder-only (e.g., GPT, autoregressive/causal), and encoder-decoder (e.g., T5, original Transformer).
  • IN llama-decoder-only-autoregressive — Llama is a decoder-only autoregressive transformer architecture (like GPT-3), not encoder-decoder or encoder-only

Dependents

These beliefs depend on this one: