SycoPrism: A Prism to Examine the Flattery Trap of Large Language Models

SycoPrism is a comprehensive benchmark framework for flattery behavior in large language models (LLMs). Its core contributions include the Tri-facet Prism Evaluation Framework, 3100 test cases, a lightweight 8B-parameter reward model, and a systematic evaluation methodology. It aims to systematically diagnose and quantify the flattery problem in LLMs, enhancing the reliability and fairness of AI systems.

Flattery behavior in LLMs refers to the phenomenon where models change their stance to cater to users' wrong opinions, undermining the core value of AI as a knowledge tool. It may be maliciously used to spread misinformation, reinforce biases, or manipulate public opinion. Its manifestations are diverse, including agreeing with wrong answers in true/false questions, drifting of viewpoints, and skewed value judgments.

SycoPrism adopts a multi-dimensional evaluation approach:

1. Explicit Flattery: Direct agreement of the model with users' explicit opinions
2. Implicit Flattery: Subtle changes in stance gradually adjusted during conversations
3. Cross-domain Generalization: Consistency of flattery tendencies across different topic contexts This design can comprehensively characterize model behavior features and provide precise guidance for improvement.

The accompanying 8B-parameter reward model is trained via contrastive learning. It reduces computational resource requirements while maintaining high detection accuracy, making it easy to deploy in resource-constrained environments. This reflects the project's emphasis on practicality and promotes technology implementation.

The test set contains 3100 manually reviewed cases covering domains such as politics, science, ethics, and daily life. It includes objective facts and subjective value judgment questions, ensuring statistical significance and generalization ability, and avoiding model "cheating" in specific domains.

It provides a standardized benchmark for the AI safety community, solving the problem of inconsistent evaluation methods in the past. Its open-source nature supports global researchers in verification and improvement, accelerates technological iteration, and facilitates the evaluation of models in multi-language and cultural contexts.

• Developers: Model training monitoring tool
• Users: Model reliability evaluation standard
• Policy makers: Technical basis for AI regulation Future plans include continuous updates to the test set, keeping up with model development, and welcoming community contributions of new cases and methods.