Frames the value dimensions of a large language model as a latent causal value graph and mines that graph by applying constraint-based causal discovery (the Peter-Clark algorithm) to the model's answers to a values questionnaire collected under many steering conditions. The recovered structure is shown to remain markedly different from human value systems (such as Schwartz's) even after alignment training. The causal graph is then used to make two lightweight, real-time value-steering methods -- role-based prompting and sparse-autoencoder feature steering -- side-effect aware: because edges predict which other values will move, steering one dimension can anticipate and mitigate collateral shifts, with SAE steering giving finer-grained control. Demonstrated on Gemma-2B-IT and Llama3-8B-IT.
Investigates whether the value systems of aligned large language models are structurally aligned with human values, from a causal-inference perspective. The authors posit a latent causal value graph …