Sycophancy Is Not One Thing: Causal Separation of Sycophantic Behaviors in LLMs
The study investigated whether LLM sycophancy stems from one mechanism or multiple processes by decomposing it into three distinct behaviors: sycophantic agreement, genuine agreement, and sycophantic praise. Researchers found these behaviors are encoded in separate directions in the model's latent space and can be independently controlled through steering interventions.