Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866918440431779840 |
|---|---|
| author | Xu, Ziwen Wu, Chenyan Sun, Hengyu Hong, Haiwen Wang, Mengru Yao, Yunzhi Huang, Longtao Xue, Hui Deng, Shumin Chu, Zhixuan Chen, Huajun Zhang, Ningyu |
| author_facet | Xu, Ziwen Wu, Chenyan Sun, Hengyu Hong, Haiwen Wang, Mengru Yao, Yunzhi Huang, Longtao Xue, Hui Deng, Shumin Chu, Zhixuan Chen, Huajun Zhang, Ningyu |
| contents | Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_02343 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics Xu, Ziwen Wu, Chenyan Sun, Hengyu Hong, Haiwen Wang, Mengru Yao, Yunzhi Huang, Longtao Xue, Hui Deng, Shumin Chu, Zhixuan Chen, Huajun Zhang, Ningyu Computation and Language Artificial Intelligence Computer Vision and Pattern Recognition Information Retrieval Machine Learning Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md. |
| title | Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics |
| topic | Computation and Language Artificial Intelligence Computer Vision and Pattern Recognition Information Retrieval Machine Learning |
| url | https://arxiv.org/abs/2602.02343 |