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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add reasoning data to pretraining | 8B hybrid model · Base model eval | average accuracy: 52.70 -> 61.05 | Improved |
| add diverse large-scale reasoning data (D_LDQ) instead of small high-quality D_SHQ in pretraining (ablation) | 8B hybrid model · Base model eval | average accuracy: 54.98 -> 64.09 | Improved |
| add D_SHQ on top of D_LDQ in pretraining (M_LMQ) (ablation) | M_LDQ · Base model eval | average accuracy: 64.09 -> 64.07 | No clear effect |
| add reasoning data to pretraining, then SFT | 8B hybrid model · SFT eval | average accuracy: 26.62 -> 35.92 | Improved |
| add reasoning data to pretraining, then SFT and RL | 8B hybrid model · RL eval | average accuracy: 37.92 -> 56.66 | Improved |
| add reasoning data to pretraining, evaluated on AIME after RL | 8B hybrid model · AIME24 after SFT+RL | pass@1: 12.29 -> 45.21 | Improved |
| double the SFT data for the non-reasoning-pretrained baseline (ablation) | M_base+SFT · SFT eval | average accuracy: 26.62 -> 34.01 | Improved |
| add reasoning pretraining (D_SHQ) versus doubling SFT data for baseline | M_base+SFT (2x) · SFT eval | average accuracy: 34.01 -> 37.33 | Improved |
| add D_SHQ to D_LDQ in pretraining before SFT on D_SHQ (ablation) | M_LDQ+SFT · SFT eval | average accuracy: 46.70 -> 50.95 | Improved |
| use small high-quality D_SHQ instead of D_LDQ for SFT (ablation) | M_res+SFT · SFT eval | average accuracy: 31.54 -> 44.99 | Improved |
| double mixed-quality D_LDQ in SFT (ablation) | M_LDQ+SFT · SFT eval | math accuracy: 28.38 -> 23.46 | Hurt |
| double mixed-quality D_LDQ in SFT (ablation) | M_LDQ+SFT · SFT eval | average accuracy: 32.84 -> 32.99 | No clear effect |
Closest papers by meaning
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.93)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.91)
- Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models (ICLR 2025 · similarity 0.91)
- Unveiling the Impact of Coding Data Instruction Fine-Tuning on Large Language Models Reasoning (AAAI 2025 · similarity 0.90)
- PRISM: Demystifying Retention and Interaction in Mid-Training (ICML 2026 · similarity 0.90)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.90)
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.
