AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Front-Loading Reasoning: The Synergy between Pretraining and Post-Training Data

Syeda Nahida Akter, Shrimai Prabhumoye, Eric Nyberg, Mostofa Patwary, Mohammad Shoeybi, Yejin Choi, Bryan Catanzaro

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 24 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 reasoning data to pretraining8B hybrid model · Base model evalaverage accuracy: 52.70 -> 61.05Improved
add diverse large-scale reasoning data (D_LDQ) instead of small high-quality D_SHQ in pretraining (ablation)8B hybrid model · Base model evalaverage accuracy: 54.98 -> 64.09Improved
add D_SHQ on top of D_LDQ in pretraining (M_LMQ) (ablation)M_LDQ · Base model evalaverage accuracy: 64.09 -> 64.07No clear effect
add reasoning data to pretraining, then SFT8B hybrid model · SFT evalaverage accuracy: 26.62 -> 35.92Improved
add reasoning data to pretraining, then SFT and RL8B hybrid model · RL evalaverage accuracy: 37.92 -> 56.66Improved
add reasoning data to pretraining, evaluated on AIME after RL8B hybrid model · AIME24 after SFT+RLpass@1: 12.29 -> 45.21Improved
double the SFT data for the non-reasoning-pretrained baseline (ablation)M_base+SFT · SFT evalaverage accuracy: 26.62 -> 34.01Improved
add reasoning pretraining (D_SHQ) versus doubling SFT data for baselineM_base+SFT (2x) · SFT evalaverage accuracy: 34.01 -> 37.33Improved
add D_SHQ to D_LDQ in pretraining before SFT on D_SHQ (ablation)M_LDQ+SFT · SFT evalaverage accuracy: 46.70 -> 50.95Improved
use small high-quality D_SHQ instead of D_LDQ for SFT (ablation)M_res+SFT · SFT evalaverage accuracy: 31.54 -> 44.99Improved
double mixed-quality D_LDQ in SFT (ablation)M_LDQ+SFT · SFT evalmath accuracy: 28.38 -> 23.46Hurt
double mixed-quality D_LDQ in SFT (ablation)M_LDQ+SFT · SFT evalaverage accuracy: 32.84 -> 32.99No clear effect

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