Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2510.10205 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915625270509568 |
|---|---|
| author | Yu, Manjiang Li, Hongji Singh, Priyanka Li, Xue Wang, Di Hu, Lijie |
| author_facet | Yu, Manjiang Li, Hongji Singh, Priyanka Li, Xue Wang, Di Hu, Lijie |
| contents | Reliable behavior control is central to deploying large language models (LLMs) on the web. Activation steering offers a tuning-free route to align attributes (e.g., truthfulness) that ensure trustworthy generation. Prevailing approaches rely on coarse heuristics and lack a principled account of where to steer and how strongly to intervene. To this end, we propose Position-wise Injection with eXact Estimated Levels (PIXEL), a position-wise activation steering framework that, in contrast to prior work, learns a property-aligned subspace from dual views (tail-averaged and end-token) and selects intervention strength via a constrained geometric objective with a closed-form solution, thereby adapting to token-level sensitivity without global hyperparameter tuning. PIXEL further performs sample-level orthogonal residual calibration to refine the global attribute direction and employs a lightweight position-scanning routine to identify receptive injection sites. We additionally provide representation-level guarantees for the minimal-intervention rule, supporting reliable alignment. Across diverse models and evaluation paradigms, PIXEL consistently improves attribute alignment while preserving model general capabilities, offering a practical and principled method for LLMs' controllable generation. Our code is available at https://github.com/V1centNevwake/PIXEL-Adaptive-Steering |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2510_10205 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | PIXEL: Adaptive Steering Via Position-wise Injection with eXact Estimated Levels under Subspace Calibration Yu, Manjiang Li, Hongji Singh, Priyanka Li, Xue Wang, Di Hu, Lijie Artificial Intelligence Reliable behavior control is central to deploying large language models (LLMs) on the web. Activation steering offers a tuning-free route to align attributes (e.g., truthfulness) that ensure trustworthy generation. Prevailing approaches rely on coarse heuristics and lack a principled account of where to steer and how strongly to intervene. To this end, we propose Position-wise Injection with eXact Estimated Levels (PIXEL), a position-wise activation steering framework that, in contrast to prior work, learns a property-aligned subspace from dual views (tail-averaged and end-token) and selects intervention strength via a constrained geometric objective with a closed-form solution, thereby adapting to token-level sensitivity without global hyperparameter tuning. PIXEL further performs sample-level orthogonal residual calibration to refine the global attribute direction and employs a lightweight position-scanning routine to identify receptive injection sites. We additionally provide representation-level guarantees for the minimal-intervention rule, supporting reliable alignment. Across diverse models and evaluation paradigms, PIXEL consistently improves attribute alignment while preserving model general capabilities, offering a practical and principled method for LLMs' controllable generation. Our code is available at https://github.com/V1centNevwake/PIXEL-Adaptive-Steering |
| title | PIXEL: Adaptive Steering Via Position-wise Injection with eXact Estimated Levels under Subspace Calibration |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2510.10205 |