s1ghhh/VLADrop_LingbotVLA_LIBERO_Drop18Block_Uniform

VLADrop-LingbotVLA-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: Lingbot-VLA Drop Half (18/36 dropped)
Dropped blocks	VLM (Qwen2.5-VL) blocks [1,3,5,...,35] (all odd-indexed; attn+mlp)
Recovery training	batch size 16, 50K steps, lr 5e-5
LIBERO success rate	Spatial 85.6 / Object 97.0 / Goal 84.2 / Long 67.8 / Avg 83.7

Usage

Lingbot-VLA HF-format checkpoint. Use with the VLADrop lingbot-vla code (https://github.com/s1ghhh/VLADrop): serve with python -m deploy.lingbot_libero_policy --model_path <this_repo_local_path> --port 8200 --use_length 8 then run experiment/libero/libero/run_libero_eval.py --model_family instruct_vla. Requires the Qwen2.5-VL-3B-Instruct tokenizer (env QWEN25_PATH). The drop lists are recorded in the model config produced at export time.

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}
}