AI & Experiments / 2026

Personal Values Modeling

Testing whether lightweight model adaptation can predict personal preferences from questionnaire responses.

The goal

Explore how prompting and lightweight adaptation affect a language model’s predictions of an individual’s preferences.

Compared prompting and LoRA-based adaptation using questionnaire responses.
Implemented a residual linear head with approximately 4,096 trainable parameters.
Evaluated prediction accuracy against an unadapted baseline and repeated tests across models.

Initial accuracy gains did not reproduce consistently across models. The experiments do not establish a reliable improvement over the baseline.

An experimental project, not a validated personalization system. The reproducibility limitation is part of the result.