Pith Number

pith:3WYX72X5

pith:2024:3WYX72X5SQT2EY5MTNNX24CJYY

not attested not anchored not stored refs resolved

Better & Faster Large Language Models via Multi-token Prediction

Badr Youbi Idrissi, Baptiste Rozi\`ere, David Lopez-Paz, Fabian Gloeckle, Gabriel Synnaeve

Training language models to predict multiple future tokens improves coding performance and speeds up inference

arxiv:2404.19737 v1 · 2024-04-30 · cs.CL

Open paper page JSON Open Graph Bundle Merged state Verified badge What is a Pith Number?

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{3WYX72X5SQT2EY5MTNNX24CJYY}

Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge

Record completeness

1 Bitcoin timestamp

2 Internet Archive

3 Author claim open · sign in to claim

4 Citations open

5 Replications open

✓ Portable graph bundle live · download bundle · merged state

The bundle contains the canonical record plus signed events. A mirror can host it anywhere and recompute the same current state with the deterministic merge algorithm.

Claims

C1strongest claim

Our 13B parameter models solves 12 % more problems on HumanEval and 17 % more on MBPP than comparable next-token models. ... models trained with 4-token prediction are up to 3 times faster at inference, even with large batch sizes.

C2weakest assumption

That the reported gains are caused by the multi-token auxiliary objective rather than differences in hyper-parameters, data ordering, or other uncontrolled training details, and that the benefit persists without degradation at much larger scales.

C3one line summary

Multi-token prediction training yields higher sample efficiency, better benchmark scores on code generation, and up to 3x faster inference than standard next-token prediction for LLMs.

References

23 extracted · 23 resolved · 5 Pith anchors

[1] Program Synthesis with Large Language Models · arXiv:2108.07732

[2] Evaluating Large Language Models Trained on Code · arXiv:2107.03374

[3] Training Verifiers to Solve Math Word Problems · arXiv:2110.14168

[4] High Fidelity Neural Audio Compression · arXiv:2210.13438

[5] Leveraging parsbert and pretrained mt5 for persian abstractive text summarization 2021

Formal links

2 machine-checked theorem links

Cited by

29 papers in Pith

TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection

FastTab: A Fast Table Recognizer with a Tiny Recursive Module and 1D Transformers

Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design

Draft Less, Retrieve More: Hybrid Tree Construction for Speculative Decoding

Covert Multi-bit LLM Watermarking: An Information Theory and Coding Approach

Receipt and verification

First computed	2026-05-17T23:38:47.884145Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

ddb17feafd9427a263ac9b5b7d7049c62847350bb03b79d8d8bae0146738cb33

Aliases

arxiv: 2404.19737 · arxiv_version: 2404.19737v1 · doi: 10.48550/arxiv.2404.19737 · pith_short_12: 3WYX72X5SQT2 · pith_short_16: 3WYX72X5SQT2EY5M · pith_short_8: 3WYX72X5

Agent API

Resolver JSON Graph JSON Events JSON Schema Signing key

Verify this Pith Number yourself

curl -sH 'Accept: application/ld+json' https://pith.science/pith/3WYX72X5SQT2EY5MTNNX24CJYY \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: ddb17feafd9427a263ac9b5b7d7049c62847350bb03b79d8d8bae0146738cb33

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "416350975cb8b534fbb7c32d3860af343ee9d2b1896c0f7273bab7be92c7a500",
    "cross_cats_sorted": [],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.CL",
    "submitted_at": "2024-04-30T17:33:57Z",
    "title_canon_sha256": "193f5844a8c69bc27a3c9495377ee7c2f62d7a994b0bf3c8691d4309735d0477"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2404.19737",
    "kind": "arxiv",
    "version": 1
  }
}