Towards Steering without Sacrifice: Principled Training of Steering Vectors for Prompt-only Interventions
Paper β’ 2605.05983 β’ Published β’ 1
OpenReview: https://openreview.net/forum?id=AaT3liS5PE
Paper: https://arxiv.org/abs/2605.05983
Data: https://huggingface.co/datasets/colored-dye/concept500-contrastive
Setups:
2b_l10: 10th layer of google/gemma-2-2b-it9b_l20: 20th layer of google/gemma-2-9b-itq25_32b_l32: 32nd layer of qwen/Qwen2.5-32B-InstructDirectory structure:
.
βββ 2b_l10 -- setup
β βββ outputs_add_free
β βββ all -- full-sequence intervention
β β βββ lang ---- Lang. objective
β β β βββ 0 ------ concept 0
β β β βββ 1 ------ concept 1
β β β ...
β β βββ simpo -- SimPO objective
β βββ f2+l2 ---- prompt-only intervention (2 prefix tokens, 2 suffix tokens)
β βββ lang
β βββ simpo
...
If you find our work useful, please cite:
@inproceedings{bao2026towards,
title = {Towards Steering without Sacrifice: Principled Training of Steering Vectors for Prompt-only Interventions},
author = {Bao, Yuntai and Li, Qinfeng and Yu, Xinyan and Zhang, Xuhong and Su, Ge and Zhang, Wenqi and Yan, Liu and Weng, Haiqin and Yin, Jianwei},
booktitle = {Forty-third International Conference on Machine Learning},
year = {2026},
url = {https://openreview.net/forum?id=AaT3liS5PE},
}