This story on HackerNoon has a decentralized backup on Sia.
Transaction ID: P1OwgmETyLGeDByqjAEHJ745nncqu2c_ULZhmJorEts
Cover

Strategic LLM Training: Multi-Token Prediction's Data Efficiency in Mathematical Reasoning

Written by @cosmological | Published on 2025/7/23
TL;DR —
This figure illustrates the profound impact of training scale on multi-token prediction models' performance on GSM8K, highlighting critical data efficiency considerations for mathematical reasoning.

Abstract and 1. Introduction

2. Method

3. Experiments on real data

4. Ablations on synthetic data

5. Why does it work? Some speculation

6. Related work

7. Conclusion, Impact statement, Environmental impact, Acknowledgements and References

A. Additional results on self-speculative decoding

B. Alternative architectures

C. Training speeds

D. Finetuning

E. Additional results on model scaling behavior

F. Details on CodeContests finetuning

G. Additional results on natural language benchmarks

H. Additional results on abstractive text summarization

I. Additional results on mathematical reasoning in natural language

J. Additional results on induction learning

K. Additional results on algorithmic reasoning

L. Additional intuitions on multi-token prediction

M. Training hyperparameters

I. Additional results on mathematical reasoning in natural language

Figure S13: Performance on the mathematical reasoning benchmark GSM8K (Cobbe et al., 2021). We evaluate pretrained next-token and multi-token prediction models trained on 200B and 500B tokens of natural language in 8-shot mode using nucleus sampling (Holtzman et al., 2020) with probability mass 0.95 and various sampling temperatures. Reported are the frequencies of the correct final answer to appear among k samples, for k = 1, 10, 100, estimated from 200 samples like in code generation benchmarks (Chen et al., 2021). After 200B tokens, the 2-token prediction model has a clear advantage over the next-token baseline but the order reverses after 500B tokens. The 4-token prediction model is worse throughout. We interpret this similarly to the findings in Section 4.1: the follow-your-nose chains-of-thought required for GSM8K may be difficult to learn from a limited amount of data, attesting to the data efficiency of multi-token prediction training. Once the correct circuits for correct autoregressive chains-of-thought in this domain have formed, however, multi-token prediction comes at a cost.

Authors:

(1) Fabian Gloeckle, FAIR at Meta, CERMICS Ecole des Ponts ParisTech and Equal contribution;

(2) Badr Youbi Idrissi, FAIR at Meta, LISN Université Paris-Saclayand and Equal contribution;

(3) Baptiste Rozière, FAIR at Meta;

(4) David Lopez-Paz, FAIR at Meta and a last author;

(5) Gabriel Synnaeve, FAIR at Meta and a last author.

This paper is available on arxiv under CC BY 4.0 DEED license.

[story continues]

Written by
@cosmological
From Big Bang's singularity to galaxies' cosmic dance the universe unfolds its majestic tapestry of space and time.
Topics and
tags
multi-token-prediction|llm-training|ai-optimization|natural-language-math|multi-token-llm|llm-performance|ai-evaluation|transformer-models
This story on HackerNoon has a decentralized backup on Sia.
Transaction ID: P1OwgmETyLGeDByqjAEHJ745nncqu2c_ULZhmJorEts