AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Reformulation for Pretraining Data Augmentation

Xintong Hao, Rui-Jie Zhu, Ge Zhang, Ke Shen, Chenggang Li

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 17 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add MGA reformulated data to the pretraining mix134M model · LightEval averageaverage accuracy: 31.51 -> 31.77Improved
add MGA reformulated data to the pretraining mix377M model · LightEval averageaverage accuracy: 34.57 -> 35.52Improved
add MGA reformulated data to the pretraining mix1.7B model · LightEval averageaverage accuracy: 41.15 -> 43.4Improved
add MGA reformulated data to the pretraining mix1.7B model · TriviaQAaccuracy: 4.95 -> 20.42Improved
expand 50B data with 200B MGA tokens instead of 10x repetition13B model, EntireSet · benchmark average at 200-500B tokensaverage accuracy: +2.65/+3.14/+3.43/+3.46Improved
collect more unique data (Full-Fineweb-Edu 195B) instead of repeating 50B13B model, EntireSet · benchmark average at 200-500B tokensaverage accuracy: +0.2/+0.15/-0.16/+0.11No clear effect
add MGA expansion instead of upsampling high-quality portion377M to 13B, Subset · benchmark averageaverage accuracy: +1.46/+2.67/+3.59/+3.73 vs +0.89/+1.53/+1.23/+1.41Improved
use MGA instead of Cosmopedia as synthetic source377M model · synthetic data comparisonaverage accuracy: 35.57 -> 37.28Improved
use Nemotron diverse QA pairs instead of MGA377M model · synthetic data comparisonaverage accuracy: 37.34 -> 40.72Hurt
replace all real fineweb-edu data with MGACorpus (MGA-Only) (ablation)1.7B model · LightEval averageaverage accuracy: 43.40 -> 42.45Hurt
use strict information-preserving prompt (SLM-Strict) instead of base (ablation)MGA reformulator · high-repetition pretrainingvalidation loss scaling: 78.37% vs 71.06% score>=4Mixed
use relaxed prompt (SLM-Relaxed) instead of base (ablation)MGA reformulator · high-repetition pretrainingbenchmark performance: 13.63% vs 71.06% score>=4Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy