π€ AI Summary
This study addresses the lack of bidirectional monotonic controllability in large language model sycophancy by proposing the PAS framework. This method introduces a novel activation steering mechanism based on PCA subspace decomposition and asymmetric scaling exponents, enabling precise bidirectional regulation of sycophancy through residual stream analysis. Experimental results demonstrate that PAS achieves a monotonicity score of 0.92 with an average adjustment magnitude of 15.4%, significantly outperforming existing baselines. By effectively resolving the non-monotonicity challenge inherent in sycophancy modulation, this work provides a more granular and reliable controllability solution for aligning large language models.
π Abstract
Large language models (LLMs) exhibit sycophancy, a tendency to agree with user beliefs regardless of factual accuracy. This can reinforce misconceptions, but eliminating it entirely risks over-correction against valid opinions. Effective control must therefore both reduce and increase sycophancy with predictable and gradual effect. Yet, existing methods fail to ensure a bidirectional and monotonic relationship between steering strength and behavioral outcome across models and datasets. We introduce PCA-guided Activation Scaling (PAS), an activation steering framework that decomposes residual stream activations into a PCA-identified sycophancy-honesty subspace and an orthogonal residual, then applies distinct scaling exponents to achieve monotonic, bidirectional control. Across three LLMs and three datasets, PAS achieves strong monotonicity (Spearman $Ο$ = +0.92) and an average shift of 15.4% per direction, compared with 8.7% for the baselines. Ablation studies confirm that the decomposition, asymmetric exponents, and layer selection are each essential for maintaining monotonic control. The data and code are available at https://github.com/Bellafc/PCS.