Reinforcement learning from human feedback (RLHF) is a popular technique for training high-quality AI assistants. However, RLHF may also encourage model responses that match user beliefs over truthful responses, a behavior known as sycophancy. We investigate the prevalence of sycophancy in RLHF-trained models and whether human preference judgments are responsible. We first demonstrate that five state-of-the-art AI assistants consistently exhibit sycophancy behavior across four varied free-form text-generation tasks. To understand if human preferences drive this broadly observed behavior of RLHF models, we analyze existing human preference data. We find that when a response matches a user's views, it is more likely to be preferred. Moreover, both humans and preference models (PMs) prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time. Optimizing model outputs against PMs also sometimes sacrifices truthfulness in favor of sycophancy. Overall, our results indicate that sycophancy is a general behavior of RLHF models, likely driven in part by human preference judgments favoring sycophantic responses.

Software developers increasingly rely on automated methods to assess the correctness of their code. One such method is property-based testing (PBT), wherein a test harness generates hundreds or thousands of inputs and checks the outputs of the program on those inputs using parametric properties. Though powerful, PBT induces a sizable gulf of evaluation: developers need to put in nontrivial effort to understand how well the different test inputs exercise the software under test. To bridge this gulf, we propose Tyche, a user interface that supports sensemaking around the effectiveness of property-based tests. Guided by a formative design exploration, our design of Tyche supports developers with interactive, configurable views of test behavior with tight integrations into modern developer testing workflow. These views help developers explore global testing behavior and individual test inputs alike. To accelerate the development of powerful, interactive PBT tools, we define a standard for PBT test reporting and integrate it with a widely used PBT library. A self-guided online usability study revealed that Tyche’s visualizations help developers to more accurately assess software testing effectiveness.