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

Researchers at arXiv introduced a dual-stance evaluation method to test whether activation steering on Llama-3-8B-Instruct reduces sycophancy without suppressing agreement with factually correct statements. The study found that the steering direction projected equally onto both sycophantic and factual agreement subspaces, reducing agreement with correct statements such as "the Earth is round" alongside sycophantic ones. The findings reveal a general limitation: representations readable from activations may not be writable through them.

arXiv:2606.11205v1 Announce Type: new Abstract: Activation steering can shift LLM behaviour, but standard evaluations do not typically test whether a sycophancy-reduction direction also suppresses agreement with factually correct statements. We introduce dual-stance evaluation, which tests both stances of each topic, and apply it to centroid-difference steering on Llama-3-8B-Instruct. We find a dissociation: the model represents sycophantic and factual agreement in geometrically distinct subspaces, yet the steering direction projects equally onto both and cannot differentially target either. The direction accordingly reduces agreement with factually correct statements e.g. that the Earth is round as well as sycophantic ones. All other static properties of the two activation groups are matched, suggesting the behavioural dissociation arises from generation dynamics or from finer-grained structure that residual-stream analysis cannot resolve. The pattern illustrates a general gap: representations that are readable from activations may not be writable through them.