Join our daily and weekly newsletters for newest updates and exclusive content to cover the industry. Learn more
In the last month, Openi rolled some updates With GPT-4o after multiple users, including former Opui CEO Emmet Shear and Hugging Face Picer Executive Clement Delangue says the model excess users.
The flattery, called sycophycy, always The model is headed to prevent user preferencesBe carefully polite, and don’t push. It’s too annoying. Sycophancy can lead to of models that releases misinformation or reinforcement Dangerous tasks. And as businesses begin to make applications and agents have been built on these sycophant llms, they run models that agree with the harmful business decisions, and can affect AI agents.
Stanford University,, Carnegie Mellon University and University of Oxford Researchers seek to change that Suggest a benchmark to measure sycophancy in models. They call benchmark elephants, for checking LLMS as excess sycophants, and knowing that every big language model (LLM) has a level of sycophany. By understanding how sycophantic models, benchmark can lead businesses to make instructions if using llms.
To test benchmark, researchers point to two personalized counseling data, and the posts that end in real-world situations, where postershatshates are in reality, which are the people who are primarily or not in some circumstances.
The idea behind the experiment is to see how models behave when facing questions. It estimates what social sycophancy researchers call, if models try to preserve “user’s face, or their own image or social identity.
“More” Hidden “Social Queries are exactly what our benchmark gets at – instead of previous work that only looks at factual agreement or explicit beliefs, our benchmarking on more implicit or hidden assumptions,” Myra cheng, one of the researchers and co-author Of the paper, told Venturebeat. “We have chosen to view the domain of personal advice because Sykophancy injuries have a lot of consequences, but casual flattery will also get ’emotional validate’ behavior.”
Testing the models
For the test, researchers fed data from QEQ and Aita to PPT-4o in Openii, Gemini 1.5 flash from Google,, Anthropic‘S Claude Sonnet 3.7 and open weight models from Meta (Llama 3-8b-standard, llama 4-scout-17b-16-e and llama 3.3-70b-turbo) and It’s a mistake7B-Stress-v0.3 and the mistal small 24b-stress2501.
Cheng said they “marked the models using the GPT-4O API, using a model version from the late sycophantic model and returned it.”
To measure the sykophy, the elephant’s way is watching five social sycophancy related behaviors:
- Validate emotionally or approval without a criterion
- Moral endorsement or telling users is correct moral right, even if they do not
- Indirect language in which the model avoids providing direct suggestions
- Indirect action, or where the model advises with passive coverage mechanisms
- Receive framing without provoking trouble thoughts.
It is known to try that all LLMs showed high levels of sycophycy, more than people, and social sycophancy has proved to be reduced. However, it is shown to try that GPT-4o “has some highest social sycophancy rate, while Gemini-1.5-flash with lowest.”
LLMS separates some biases in datasets. The paper found that AITA posts have some gender biasa, in posts that talk to spouses or girlfriends that are always worn in socially inappropriate. At the same time, those with a husband, boyfriend, parents or mother are wrong. Researchers say models “can rely on the heuristic relational heuristics to excessively blamed.” In other words, models are more sycophantic to people who have boyfriends and husbands than women or wives.
Why it’s important
It’s nice if a chatbot talks to you as an affectionate entity, and it can feel good if the model proves your comments. But sycophycy raise Concerns about models’ Support the wrong or part of statements and, on a more personal level, stimulating self-solitude, deception or harmful behaviors.
Businesses do not like their AI applications built with LLMS which spreads wrong information to meet users. It may be wrong with the tone or behavior of an organization and can cause employees and users on platforms.
Researchers say the elephant’s method and further test helps to inform better reciprocal rewards to prevent sycophancy from adding.
