Join our daily and weekly newsletters to obtain the latest updates and exclusive content on the coverage of the industry leader. Get more information
Last month, Operai backed some updates to GPT-4O after several users, including former OpenI CEO, Emmet Shear and Hugging Face executive, Clement Delangue, said the model surpassed flattering users.
The adulation, called Sicofanance, often led the model to differ to the user’s preferences, be extremely educated and not go back. It was also upset. Sycophanancy could lead to models to free Missions to harmful behaviors. And as companies begin to make applications and agents built on these LLMS Sycophant, they work with the risk that models accept harmful commercial decisions, encouraging false information to extend and be used by AI agents, and Mayy.
The University of Stanford, the Carnegie Mellon University and the researchers at the University of Oxford sought to change that by proposing a reference point to measure the skotfancia of the models. They called the reference elephant, for the evaluation of LLM as excessive systems, and discovered that each large language model (LLM) has a certain level of Sycophany. By understanding how sycophanical models can be, the reference point can guide companies about the creation of guidelines when using LLM.
To prove the reference point, the researchers pointed to the models to two sets of personal advice data: the QEQ, a set of open personal advice questions about real -world and AITA sitations, publications of the Subeddit R/Amoithashole in Judthasshole in Judthashole in judge and committee some situations.
The idea behind the experiment is to see how the models behave when they face consultations. Evaluate what the researchers called Social Sycofanance, if the models try to preserve the “face” of the user, or social identity or social identity.
“More social consultations” hidden “are exactly what our reference point obtains in the previous work that only analyzes the de facto agreement or explicit beliefs, our reference point captures the agreement or adulation based on more imports and co -documents, said Venturebeat.”
Models test
For the test, The Refrachers Fed The Data From qeq and aita to Openai’s GPT-4O, Gemini 1.5 Flash from Google, Anthropic’s Claude Sonnet 3.7 and Open Weight Models From Meta and call and call and call and call and call and call and flame and call-7b-16b-16b-16b-16-e and calls 7b-instruct-v0.3 and the small abuse-24b-instruct2501.
Cheng said that “they compared the models using the API GPT-4O, which uses a version of the model at the end of 2024, before both OpenAI implemented the new Sycophanic model and returned back.”
To measure psychophanship, the elephant method analyzes five behaviors related to social sycophanancy:
- Emotional validation or overstress without criticism
- Moral support or say that users are morally right, even when they are not
- Indirect language where the model avoids giving direct suggestions
- Indirect action, or where model advisors with passive coping mechanisms
- Couple acceptance that does not defy problematic assumptions.
The test found that all LLMs showed high levels of sycophanancy, equally more than humans, and social sycophycia proved to be diffusion to mitigate. However, the test showed that GPT-4O “has some of the highest social sicofanance rates, while Gemini-1.5-Flash has the lowest.”
The LLM also amplified some biases in the data sets. The document indicated that Aita publications had a gender bias, since the publications mentioned by wives or brides were more or correctly marked as socially inappropriate. At the same time, those with husband, boyfriend, father or mother were poorly classified. The researchers said the models “can depend on the relational heuristics of gender in the fault of overload and little bird.” In other words, the models were more sycopophant for people with bride and groom and bands than for those with girlfriends or wives.
Why is it important
It is good if a chatbot speaks to him as an empathic entity, and he can feel if the model validates his comments. But the Sileno Conerns raises on the models that support false or worrying statements and, at a more personal level, they could foster selfisolation, delusions. or harmful behavior.
Companies do not want their AI applications created with LLM that disseminate false information be aggravated for users. It can be misaligned with a tone or ethics of an organization and could be very annoying for employees and end users of their platforms.
The researchers said that the elephant method and additional tests could help inform better railings to prevent sycophanancy from increasing.
