s1ghhh/VLADrop_GigaBrain0_LIBERO_Drop13Block_Uniform

VLADrop-GigaBrain0-LIBERO-drop-half

Checkpoint for Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?.

DTR (Drop-Then-Recovery) removes transformer blocks from a pretrained VLA model and recovery-fine-tunes the smaller dense model. Code: https://github.com/s1ghhh/VLADrop

This checkpoint

Paper row	Table 7: GigaBrain-0 Drop Half (13/26 dropped)
Dropped blocks	VLM (PaliGemma2) blocks [1,3,5,...,25] (all odd-indexed; attn+mlp; pass --llm_drop_attn_list/--llm_drop_mlp_list at load time)
Recovery training	batch size 16, 50K steps, lr 2.5e-5
LIBERO success rate	Spatial 85.0 / Object 98.6 / Goal 93.2 / Long 75.2 / Avg 88.0

Usage

GigaBrain-0 checkpoint (model/ format of the giga_models package). Use with the VLADrop giga-brain-0 eval code (https://github.com/s1ghhh/VLADrop):

python 0_eval/run_libero_eval.py \
    --model_path <this_repo_local_path> \
    --norm_stats_path norm_stats_gigabrain.json \
    --task_suite_name libero_spatial --num_trials_per_task 50 --replan_steps 5 \
    [--llm_drop_attn_list ... --llm_drop_mlp_list ...]

Important: for dropped variants, pass the exact drop lists shown above at load time (the drop is applied by a runtime patch, not stored in the weights). norm_stats_gigabrain.json (LIBERO norm stats) is included in this repo.

Citation

@article{sun2026vladrop,
  title={Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?},
  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},
  journal={arXiv preprint arXiv:2606.27755},
  year={2026}
}