CGL studies continual learning for GUI agents: adapting to new application domains while retaining previously acquired interaction skills. It brings together the adaptation efficiency of supervised fine-tuning and the retention behavior of reinforcement learning through three components:
- Error-Aware Routing: introduce corrected-response supervision when a rollout group fails to reach the full action reward.
- Entropy-Regulated Tuning: adjust the supervised objective's weight using policy entropy.
- Conditional Gradient Surgery: project conflicting supervised gradients relative to the GRPO update.
We also introduce AndroidControl-CL, a benchmark that organizes Android GUI interactions into seven application categories for sequential training and evaluation.
- [2026.10.16] CGL training code and launch scripts are available. Access AndroidControl-CL on Hugging Face.
- [CVPR 2026] Read the conference paper.
The AndroidControl-CL dataset includes screenshots, task annotations, and category-specific training/test files.
| Task category | Parquet filename prefix |
|---|---|
| Shopping | shopping |
| Productivity and Office | productivity |
| Communication | communication |
| Travel and Transportation | travel_and_local |
| System Tools | tools |
| Education and Science | education |
| Life and Entertainment | entertainment |
Download the files for the current task from parquet_convert/, for example productivity_train.parquet and productivity_test.parquet. See training notes for the input schema and supported actions.
The original experiments use 8 Ascend NPUs and vLLM 0.7.0. Prepare the matching CANN, PyTorch/torch-npu, and vllm-ascend environment, then install:
git clone https://github.com/skounputer/CGL.git
cd CGL
pip install -r requirements-npu.txt
pip install tensorboard
pip install -e . --no-depsThis repository includes the verl 0.4.1.dev backbone. The Ascend installation uses requirements-npu.txt; the upstream [vllm] extra is intended for CUDA installations. Additional backend details are in the training notes.
Set the current task's data and the initial or previous-task Hugging Face checkpoint:
export TRAIN_FILES=/path/to/productivity_train.parquet
export VAL_FILES=/path/to/productivity_test.parquet
export MODEL_PATH=/path/to/previous_task_hf_checkpoint
bash scripts/train_cgl.shUse DRY_RUN=1 bash scripts/train_cgl.sh to inspect the command. For continual learning, train one task, export its actor checkpoint, and pass the resulting Hugging Face model as MODEL_PATH for the next task.
| Setting | Override passed to the training script |
|---|---|
| CGL | Default launcher configuration |
| Direct gradient addition | actor_rollout_ref.actor.sft_grpo_loss_add_linear=True |
| Fixed supervised weight | actor_rollout_ref.actor.sft_grpo_mu=False |
| GRPO only | actor_rollout_ref.actor.sft_grpo_loss=False |
bash scripts/train_cgl.sh actor_rollout_ref.actor.sft_grpo_mu=FalseThe CPU checks cover CGL's routing, loss, gradient update, reward callback, and configuration:
pip install -r requirements-cgl-test.txt
pytest -q tests/cglFull NPU training revalidation is pending for this release. Initial SFT, model weights, and the complete evaluation pipeline are not included. See validation details and implementation changes.
If you use CGL or AndroidControl-CL, please cite:
@InProceedings{Yao_2026_CVPR,
author={Yao, Zhenquan and Huang, Zitong and Zeng, Yihan and Han, Jianhua and Xu, Hang and Feng, Chun-Mei and Ma, Jianwei and Zuo, Wangmeng},
title={CGL: Advancing Continual GUI Learning via Reinforcement Fine-Tuning},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month={June},
year={2026},
pages={15212-15221},
url={https://openaccess.thecvf.com/content/CVPR2026/html/Yao_CGL_Advancing_Continual_GUI_Learning_via_Reinforcement_Fine-Tuning_CVPR_2026_paper.html}
}Our training infrastructure builds on verl / HybridFlow. The code is released under Apache-2.0, with upstream notices retained. Dataset and model licenses apply separately.