Survey-to-Behavior: Downstream Alignment of Human Values in LLMs via Survey Questions
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_ | 1866909737863348224 |
|---|---|
| author | Nie, Shangrui Mai, Florian Kaczér, David Welch, Charles Zhao, Zhixue Flek, Lucie |
| author_facet | Nie, Shangrui Mai, Florian Kaczér, David Welch, Charles Zhao, Zhixue Flek, Lucie |
| contents | Large language models implicitly encode preferences over human values, yet steering them often requires large training data. In this work, we investigate a simple approach: Can we reliably modify a model's value system in downstream behavior by training it to answer value survey questions accordingly? We first construct value profiles of several open-source LLMs by asking them to rate a series of value-related descriptions spanning 20 distinct human values, which we use as a baseline for subsequent experiments. We then investigate whether the value system of a model can be governed by fine-tuning on the value surveys. We evaluate the effect of finetuning on the model's behavior in two ways; first, we assess how answers change on in-domain, held-out survey questions. Second, we evaluate whether the model's behavior changes in out-of-domain settings (situational scenarios). To this end, we construct a contextualized moral judgment dataset based on Reddit posts and evaluate changes in the model's behavior in text-based adventure games. We demonstrate that our simple approach can not only change the model's answers to in-domain survey questions, but also produces substantial shifts (value alignment) in implicit downstream task behavior. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2508_11414 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Survey-to-Behavior: Downstream Alignment of Human Values in LLMs via Survey Questions Nie, Shangrui Mai, Florian Kaczér, David Welch, Charles Zhao, Zhixue Flek, Lucie Computation and Language Large language models implicitly encode preferences over human values, yet steering them often requires large training data. In this work, we investigate a simple approach: Can we reliably modify a model's value system in downstream behavior by training it to answer value survey questions accordingly? We first construct value profiles of several open-source LLMs by asking them to rate a series of value-related descriptions spanning 20 distinct human values, which we use as a baseline for subsequent experiments. We then investigate whether the value system of a model can be governed by fine-tuning on the value surveys. We evaluate the effect of finetuning on the model's behavior in two ways; first, we assess how answers change on in-domain, held-out survey questions. Second, we evaluate whether the model's behavior changes in out-of-domain settings (situational scenarios). To this end, we construct a contextualized moral judgment dataset based on Reddit posts and evaluate changes in the model's behavior in text-based adventure games. We demonstrate that our simple approach can not only change the model's answers to in-domain survey questions, but also produces substantial shifts (value alignment) in implicit downstream task behavior. |
| title | Survey-to-Behavior: Downstream Alignment of Human Values in LLMs via Survey Questions |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2508.11414 |