New Method Controls LLM Sycophancy Using Cascading Linear Features

Controlling specific behaviors in large language models often relies on activation steering, which requires numerous contrastive data samples to identify the underlying features. This paper presents a novel iterative data generation pipeline designed to isolate "cascading linear features" responsible for particular model behaviors. The core idea is to move beyond simple binary sample pairs and instead use samples that exhibit varying degrees of a feature, allowing for a more precise disentanglement of the features within the model's activation space. The researchers applied this method to address sycophancy in LLMs, which is the tendency for models to overly prioritize user validation. They demonstrated that the sycophancy features identified through their cascading sample approach form linearly separable subspaces, leading to clearer detection and more effective steering. This technique not only matches or outperforms traditional LLM-as-a-judge and system prompting methods but also offers lower computational demand and enhanced interpretability.