Mechanistic Relationship Between Sycophantic Agreement and Honesty/Deception in LLMs

Ascertain the mechanistic relationship between sycophantic agreement in instruction-tuned large language models and the internal features underlying honesty and deception. Determine whether sycophantic agreement shares causal mechanisms with honesty or deception or is represented as a distinct feature, and characterize how these mechanisms interact across model layers and architectures.

Background

The paper demonstrates that sycophantic agreement, genuine agreement, and sycophantic praise are encoded along distinct, linearly separable directions in model activation space and can be independently steered without cross-effects. Geometric analyses show early entanglement of agreement signals with later divergence between genuine and sycophantic agreement, while sycophantic praise remains orthogonal.

Despite establishing separability among these sycophancy-related behaviors, the authors explicitly note that the relationship between sycophantic agreement and broader constructs—specifically honesty and deception—has not been mechanistically characterized. Prior work has probed related high-level behaviors (e.g., truthfulness, deception), but direct evidence connecting their internal representations to sycophantic agreement remains unresolved, motivating targeted mechanistic investigation.