ATHENA: Accelerated Multi-Task Heterogeneous Influence Functions for Robot Data Curation
Influence functions at billion-parameter VLA scale (about 313x faster via Kronecker structure and a rank-r Hessian approximation) for multi-task data curation; matches full-data fine-tuning with 50% of demos in sim.将影响函数扩展到十亿参数级 VLA（借助 Kronecker 结构与秩 r 的 Hessian 近似提速约 313 倍）用于多任务数据筛选；在仿真中只用 50% 示教即可达到全量数据微调的效果。
BibTeX
@misc{xu2026athena,
title = {{ATHENA: Accelerated Multi-Task Heterogeneous Influence Functions for Robot Data Curation}},
author = {Tao Xu and Jiaxin Wang and Runhao Zhang and Jiayi Guan and Xianchao Zeng and Weixi Song and Xinyu Zhou and Zhetao Chen and Guang Chen and Yong-Lu Li},
year = {2026},
eprint = {2606.16208},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2606.16208},
}