Views
No views yet
| Paper row | Table 1 'Drop Half Language' / Table 2 'Block' / Table 5 'DTR-16' (1.64x task speedup) |
| Dropped blocks | Llama-2 decoder blocks [1,3,5,...,31] (all odd-indexed; whole blocks, baked into config.json) |
| Recovery training | LoRA rank 32 (merged), batch size 16, 50K steps, lr 5e-4 |
| LIBERO success rate | Spatial 99.0 / Object 100.0 / Goal 97.8 / Long 96.4 / Avg 98.3; LIBERO-Goal speed benchmark: SR 100.0, 1.56x action speedup, 8.36 GB |
config.json (text_config.drop_attn_list / drop_mlp_list), so evaluation
reconstructs the pruned graph automatically. Requires the VLADrop modified
transformers (transformers-openvla-oft-dropped) from https://github.com/s1ghhh/VLADrop1python -m experiments.robot.libero.run_libero_eval_random_dropped \
2 --pretrained_checkpoint <this_repo_local_path> \
3 --task_suite_name libero_spatial \
4 --num_trials_per_task 50 --center_crop Trueaction_head--*.pt and proprio_projector--*.pt (L1 regression head, 2 input images + proprio).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}