Steering Risk Preferences in Large Language Models by Aligning Behavioral and Neural Representations
Fuente:
arXiv
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915291488845824 |
|---|---|
| author | Zhu, Jian-Qiao Yan, Haijiang Griffiths, Thomas L. |
| author_facet | Zhu, Jian-Qiao Yan, Haijiang Griffiths, Thomas L. |
| contents | Changing the behavior of large language models (LLMs) can be as straightforward as editing the Transformer's residual streams using appropriately constructed "steering vectors." These modifications to internal neural activations, a form of representation engineering, offer an effective and targeted means of influencing model behavior without retraining or fine-tuning the model. But how can such steering vectors be systematically identified? We propose a principled approach for uncovering steering vectors by aligning latent representations elicited through behavioral methods (specifically, Markov chain Monte Carlo with LLMs) with their neural counterparts. To evaluate this approach, we focus on extracting latent risk preferences from LLMs and steering their risk-related outputs using the aligned representations as steering vectors. We show that the resulting steering vectors successfully and reliably modulate LLM outputs in line with the targeted behavior. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2505_11615 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Steering Risk Preferences in Large Language Models by Aligning Behavioral and Neural Representations Zhu, Jian-Qiao Yan, Haijiang Griffiths, Thomas L. Computation and Language Artificial Intelligence Changing the behavior of large language models (LLMs) can be as straightforward as editing the Transformer's residual streams using appropriately constructed "steering vectors." These modifications to internal neural activations, a form of representation engineering, offer an effective and targeted means of influencing model behavior without retraining or fine-tuning the model. But how can such steering vectors be systematically identified? We propose a principled approach for uncovering steering vectors by aligning latent representations elicited through behavioral methods (specifically, Markov chain Monte Carlo with LLMs) with their neural counterparts. To evaluate this approach, we focus on extracting latent risk preferences from LLMs and steering their risk-related outputs using the aligned representations as steering vectors. We show that the resulting steering vectors successfully and reliably modulate LLM outputs in line with the targeted behavior. |
| title | Steering Risk Preferences in Large Language Models by Aligning Behavioral and Neural Representations |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2505.11615 |