Transformer Architecture · 2024

Better & Faster Large Language Models via Multi-token Prediction

Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Roziere, David Lopez-Paz, Gabriel Synnaeve (Meta)

Training with multiple parallel prediction heads improves sample efficiency and enables self-speculative decoding, and became the multi-token-prediction objective adopted by DeepSeek-V3 and later models.

Editorial record

Plain-language summary

Training with multiple parallel prediction heads improves sample efficiency and enables self-speculative decoding, and became the multi-token-prediction objective adopted by DeepSeek-V3 and later models.

Knowledge graph

Relationships

Descendants

  • Depends onEvidence: Strongly supported

    Attention Is All You Need

    Builds on the transformer architecture lineage in the archive

    freshness sweep 2026

Source record

Provenance

Record ID
P-614
Record created
2026-07-13
Last reviewed
2026-07-14
Record version
2

Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.