AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 0 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 procedural pretraining on k-Dyck sequencessmall transformer · Haystackaccuracy (%): 10 -> 98Improved
shuffle procedural sequences, destroying structure (ablation)best procedural model · Algorithmic tasksaccuracyHurt
transfer only attention layers instead of the full model (ablation)SET-pretrained model · Identity/Haystackaccuracy: 80-percentage point improvementImproved
transfer only attention layers instead of the full model (ablation)SET-pretrained model · Haystackaccuracy (%): 18.9 -> 88.9Improved
transfer only attention layers instead of the full model (ablation)SET-pretrained model · Reversed additionaccuracy (%): 44.6 -> 54.4Improved
transfer only MLP layers instead of the full model (ablation)ECA rule 110 model · Reversed additionaccuracy (%): 91.0 -> 70.5Hurt
combine SET attention with ECA MLPsSET and ECA models · Four algorithmic tasksaverage accuracy: 80.6 -> 89.3Improved
add SET attention + ECA MLP pretraining vs no pretrainingsmall transformer · Algorithmic tasksaverage accuracy: 57.4 -> 89.3Improved
add a sharp-attention regularizer instead of procedural pretraining (ablation)small transformer · Algorithmic tasksaccuracyNo clear effect
shuffle pretrained weights per layer (keep magnitudes) (ablation)best pretrained models · Algorithmic tasksaccuracyHurt
add 0.1-0.3% procedural tokens before standard pretrainingGPT-2 small · C4/CodeParrot/DeepMind-Mathloss: 2.1M tokens = 0.3%, 0.2%, 0.1%Improved
substitute procedural tokens for semantic tokensGPT-2 small · C4semantic tokens saved (%): about 45%Improved

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