Simple Synthetic Data Reduces Sycophancy In Large Language Models

Sycophancy is an undesirable behavior where models tailor their responses to follow a human user’s view even when that view is not objectively correct (e.g., adapting liberal views once a user reveals that they are liberal). In this paper, we study the prevalence of sycophancy in language models and propose a simple synthetic-data intervention to reduce this behavior. First, on a set of three sycophancy tasks (Perez et al., 2022) where models are asked for an opinion on statements with no correct answers (e.g., politics), we observe that both model scaling and instruction tuning significantly increase sycophancy for PaLM models up to 540B parameters. Second, we extend sycophancy evaluations to simple addition statements that are objectively incorrect, finding that despite knowing that these statements are wrong, language models will still agree with them if the user does as well. To reduce sycophancy, we present a straightforward synthetic-data intervention that takes public NLP tasks and encourages models to be robust to user opinions on these tasks. Adding these data in a lightweight finetuning step can significantly reduce sycophantic behavior on held-out prompts.

Introduction. Language models have seen significant advancement in recent years, including the capacity to solve complex tasks that require reasoning (Brown et al., 2020; Chowdhery et al., 2022; OpenAI, 2023; Google, 2023; Touvron et al., 2023, inter alia). As these models may one day be able to solve problems that humans cannot solve, it is important to ensure that models are aligned and avoid reward hacking (Amodei et al., 2016; Saunders et al., 2022; Bowman et al., 2022), such as exploiting the preferences of human raters (Amodei et al., 2016; Cotra, 2021). One basic form of reward hacking is sycophancy, where a model responds to a question with a user’s preferred answer in order to look favorable even if that answer is not correct (Cotra, 2021; Perez et al., 2022; Radhakrishnan et al., 2023), as shown in Figure 1. In this paper, we study sycophancy across a set of base and instruction-tuned models1 (Chowdhery et al., 2022; Chung et al., 2022, PaLM and Flan-PaLM). We then propose a straightforward syntheticdata intervention in an additional finetuning stage that reduces this behavior.

Discussion / Conclusion. Limitations. While our work sheds light on the prevalence of sycophancy and presents a simple intervention to reduce this behavior, there are several limitations to our work. First, we set our evaluations and intervention method to follow the prompt format used in Perez et al. (2022) (i.e., “Human: [question]\nAssistant:”), so it is unclear whether our results generalize to other formats that could be used. We view our findings, however, as evidence of the general potential of using straightforward synthetic data to reduce sycophancy and not as evidence that our specific set of data can solve all instances of sycophancy. Moreover, we did not conduct experimentation on correct addition statements that would verify that models can agree with correct statements (versus disagreeing with incorrect statements).