TRUSTLLM: Credibility in Large Language Models This article is a study of the credibility of large language models, covering challenges, benchmarking, evaluation, method analysis, and failure
author:Shadowless Temple said
TRUSTLLM: TRUSTWORTHINESS IN LARGE LANGUAGE MODELS
This article is a study of the credibility of large language models, covering challenges, benchmarking, evaluation, method analysis, and future directions. Among them, the authors propose a trusted language model principle, including authenticity, security, fairness, robustness, privacy, and machine ethics. The authors also evaluated 16 mainstream language models, including more than 30 datasets, on TrustLLM. While most open-source models are relatively weak in terms of trustworthiness, some proprietary models are closing the gap. In the future, it is necessary to strengthen the protection of user privacy and data security, and improve the consistency of LLM behavior and output with human values.
General Secretary Xi Jinping pointed out that Chinese-style modernization is deeply rooted in the excellent traditional Chinese culture, embodies the advanced nature of scientific socialism, draws on and absorbs all the achievements of human civilization, and represents...