Research map / Large language models / Language, safety and interpretability
Procedural Pretraining: Warming Up Language Models with Abstract Data
Liangze Jiang, Zachary Shinnick, Anton van den Hengel, Hemanth Saratchandran, Damien Teney
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 procedural pretraining on k-Dyck sequences | small transformer · Haystack | accuracy (%): 10 -> 98 | Improved |
| shuffle procedural sequences, destroying structure (ablation) | best procedural model · Algorithmic tasks | accuracy | Hurt |
| transfer only attention layers instead of the full model (ablation) | SET-pretrained model · Identity/Haystack | accuracy: 80-percentage point improvement | Improved |
| transfer only attention layers instead of the full model (ablation) | SET-pretrained model · Haystack | accuracy (%): 18.9 -> 88.9 | Improved |
| transfer only attention layers instead of the full model (ablation) | SET-pretrained model · Reversed addition | accuracy (%): 44.6 -> 54.4 | Improved |
| transfer only MLP layers instead of the full model (ablation) | ECA rule 110 model · Reversed addition | accuracy (%): 91.0 -> 70.5 | Hurt |
| combine SET attention with ECA MLPs | SET and ECA models · Four algorithmic tasks | average accuracy: 80.6 -> 89.3 | Improved |
| add SET attention + ECA MLP pretraining vs no pretraining | small transformer · Algorithmic tasks | average accuracy: 57.4 -> 89.3 | Improved |
| add a sharp-attention regularizer instead of procedural pretraining (ablation) | small transformer · Algorithmic tasks | accuracy | No clear effect |
| shuffle pretrained weights per layer (keep magnitudes) (ablation) | best pretrained models · Algorithmic tasks | accuracy | Hurt |
| add 0.1-0.3% procedural tokens before standard pretraining | GPT-2 small · C4/CodeParrot/DeepMind-Math | loss: 2.1M tokens = 0.3%, 0.2%, 0.1% | Improved |
| substitute procedural tokens for semantic tokens | GPT-2 small · C4 | semantic tokens saved (%): about 45% | Improved |
Closest papers by meaning
- Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models (ICLR 2025 · similarity 0.91)
- Exploring Visual Pretraining for Learning Language Intelligence (CVPR 2026 · similarity 0.90)
- Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages (ICLR 2025 · similarity 0.88)
- MASS: Mathematical Data Selection via Skill Graphs for Pretraining Large Language Models (ICML 2025 · similarity 0.88)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.88)
- Meaningful Learning: Enhancing Abstract Reasoning in Large Language Models via Generic Fact Guidance (NeurIPS 2024 · similarity 0.88)
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.
