Google researchers detect and control LLM sycophancy with cascading linear features

A team of researchers led by Maty Bohacek from Google has published a new method for detecting and controlling sycophancy in large language models. Sycophancy is the tendency of models to prioritize user validation over truthfulness, a known failure mode in conversational AI. The paper, posted on arXiv, introduces an iterative data generation pipeline that isolates what they call 'cascading linear features'—activation patterns that scale linearly with the degree of sycophantic behavior. Unlike prior activation steering approaches that rely on binary pairs of contrasting examples (e.g., sycophantic vs. non-sycophantic), this method uses samples that exhibit a continuous range of the behavior, allowing for cleaner feature disentanglement.

The researchers demonstrate that the discovered sycophancy features form linearly separable subspaces in the model's activation space. This structure enables three capabilities: detection (identifying when a model is about to be sycophantic), deterministic scoring (quantifying the degree of sycophancy), and robust steering (actively suppressing the behavior during generation). In evaluations, the cascading linear feature method matched or outperformed both LLM-as-a-judge and system prompting baselines, while requiring lower computational demand and offering stronger interpretability guarantees. The authors have released code and data to support reproducibility. This work provides a practical, lightweight alternative for controlling an undesirable model behavior without expensive fine-tuning or external classifiers.