This is the official code for the paper on Incremental Learning of Sparse Attention Patterns in Transformers presented at EurIPS 2025 Workshop on Principles of Generative Modeling and accepted to ICML 2026 Main Conference.
The analysis/ folder contains notebooks for regenerating the paper plots from
the W&B project r-alvarezlucendo16/incremental-learning.
uv sync# List available experiments
bash run.sh
# Run a specific experiment
bash run.sh <experiment_name>Experiments are configured using Hydra with configs located in conf/.
- Experiment configs in
conf/experiments/override base settings fromconf/train.yaml - Component configs can be customized:
model/,dataset/,optimizer/,scheduler/,loss/