Views
No views yet
| Paper row | Table 7: GigaBrain-0 baseline (0/26 dropped) |
| Dropped blocks | none (full model) |
| Recovery training | batch size 16, 50K steps, lr 2.5e-5, fine-tuned from GigaBrain-0-3.5B-Base on standard LIBERO |
| LIBERO success rate | Spatial 84.4 / Object 98.2 / Goal 93.0 / Long 76.2 / Avg 88.0 |
model/ format of the giga_models package). Use with the
VLADrop giga-brain-0 eval code (https://github.com/s1ghhh/VLADrop):1python 0_eval/run_libero_eval.py \
2 --model_path <this_repo_local_path> \
3 --norm_stats_path norm_stats_gigabrain.json \
4 --task_suite_name libero_spatial --num_trials_per_task 50 --replan_steps 5 \
5 [--llm_drop_attn_list ... --llm_drop_mlp_list ...]norm_stats_gigabrain.json (LIBERO norm stats) is included in this repo.1@article{sun2026vladrop,
2 title={Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?},
3 author={Sun, Guoheng and Feng, Kaixi and He, Shwai and Gong, Xiaochuan and He, Yexiao and Wang, Ziyao and Shen, Zheyu and Ye, Wanghao and Kompella, Ramana Rao and Liu, Gaowen and Li, Ang},
4 journal={arXiv preprint arXiv:2606.27755},
5 year={2026}
6}