Dual-Stance Evaluation of Sycophancy: The Structure of Agreement and the Limits of Intervention

The article introduces a dual-stance evaluation method to assess the impact of activation steering on LLM behavior, specifically applied to the Llama-3-8B-Instruct model. The findings reveal that while sycophantic and factual agreement are represented in distinct geometrical subspaces, the centroid-difference steering fails to differentially target them, leading to a reduction in agreement with factual statements alongside sycophantic ones. This highlights a significant limitation in current evaluation methods, indicating that observable activation patterns may not effectively translate into desired behavioral outcomes, which is crucial for practitioners aiming to fine-tune model responses.