How Value Induction Reshapes LLM Behaviour

The post details an experimental methodology for fine-tuning LLMs with specific value subsets from preference datasets. It quantifies the impact of this induction on related/contrasting value expression, safety metrics, and anthropomorphic language use through QA benchmarks. Key findings include the correlation between induced values and the expression of other values, a positive impact on safety, and an increase in sycophantic language.