Large language models (LLMs), exemplified by ChatGPT, have gained considerable attention for their excellent natural language processing capabilities. Nonetheless, these LLMs present many challenges, particularly in the realm of trustworthiness. Therefore, ensuring the trustworthiness of LLMs emerges as an important topic. This paper introduces TrustLLM, a comprehensive study of trustworthiness in LLMs, including principles for different dimensions of trustworthiness, established benchmark, evaluation, and analysis of trustworthiness for mainstream LLMs, and discussion of open challenges and future directions. Specifically, we first propose a set of principles for trustworthy LLMs that span eight different dimensions. Based on these principles, we further establish a benchmark across six dimensions including truthfulness, safety, fairness, robustness, privacy, and machine ethics. We then present a study evaluating 16 mainstream LLMs in TrustLLM, consisting of over 30 datasets. Our findings firstly show that in general trustworthiness and utility (i.e., functional effectiveness) are positively related. Secondly, our observations reveal that proprietary LLMs generally outperform most open-source counterparts in terms of trustworthiness, raising concerns about the potential risks of widely accessible open-source LLMs. However, a few open-source LLMs come very close to proprietary ones. Thirdly, it is important to note that some LLMs may be overly calibrated towards exhibiting trustworthiness, to the extent that they compromise their utility by mistakenly treating benign prompts as harmful and consequently not responding. Finally, we emphasize the importance of ensuring transparency not only in the models themselves but also in the technologies that underpin trustworthiness. Knowing the specific trustworthy technologies that have been employed is crucial for analyzing their effectiveness.
Multilingual large language models have been increasingly popular for their proficiency in comprehending and generating text across various languages. Previous research has shown that the presence of stereotypes and biases in monolingual large language models can be attributed to the nature of their training data, which is collected from humans and reflects societal biases. Multilingual language models undergo the same training procedure as monolingual ones, albeit with training data sourced from various languages. This raises the question: do stereotypes present in one social context leak across languages within the model? In our work, we first define the term ``stereotype leakage'' and propose a framework for its measurement. With this framework, we investigate how stereotypical associations leak across four languages: English, Russian, Chinese, and Hindi. To quantify the stereotype leakage, we employ an approach from social psychology, measuring stereotypes via group-trait associations. We evaluate human stereotypes and stereotypical associations manifested in multilingual large language models such as mBERT, mT5, and ChatGPT. Our findings show a noticeable leakage of positive, negative, and non-polar associations across all languages. Notably, Hindi within multilingual models appears to be the most susceptible to influence from other languages, while Chinese is the least. Additionally, ChatGPT exhibits a better alignment with human scores than other models.
Large language models (LLMs) have demonstrated impressive reasoning ability in various language-based tasks. Despite many proposed reasoning methods aimed at enhancing performance in downstream tasks, two fundamental questions persist: Does reasoning genuinely support predictions, and how reliable is the quality of reasoning? In this paper, we propose a framework \textsc{SCORE} to analyze how well LLMs can reason. Specifically, we focus on self-contradictory reasoning, where reasoning does not support the prediction. We find that LLMs often contradict themselves when performing reasoning tasks that involve contextual information and commonsense. The model may miss evidence or use shortcuts, thereby exhibiting self-contradictory behaviors. We also employ the Point-of-View (POV) method, which probes models to generate reasoning from multiple perspectives, as a diagnostic tool for further analysis. We find that though LLMs may appear to perform well in one-perspective settings, they fail to stabilize such behavior in multi-perspectives settings. Even for correct predictions, the reasoning may be messy and incomplete, and LLMs can easily be led astray from good reasoning. \textsc{SCORE}'s results underscore the lack of robustness required for trustworthy reasoning and the urgency for further research to establish best practices for a comprehensive evaluation of reasoning beyond accuracy-based metrics.
Reinforcement Learning from Human Feedback (RLHF) is a vital strategy for enhancing model safety in language models. However, annotating preference data for RLHF is a resource-intensive and creativity-demanding process, while automatic generation methods face limitations in data diversity and quality. In response, we present Safer-Instruct, a novel pipeline for semi-automatically constructing large-scale preference datasets. Our approach leverages reversed instruction tuning, instruction induction, and expert model evaluation to efficiently generate high-quality preference data without human annotators. We evaluate Safer-Instruct using LLaMA for instruction induction and GPT-4 as an expert model, generating approximately 10K preference samples. Finetuning an Alpaca model on this dataset demonstrates improved harmlessness while maintaining competitive performance on conversation and downstream tasks. Safer-Instruct addresses the challenges in preference data acquisition, advancing the development of safer and more responsible AI systems. Our code and data are available at https://github.com/uscnlp-lime/safer-instruct
People from different social and demographic groups express diverse perspectives and conflicting opinions on a broad set of topics such as product reviews, healthcare, law, and politics. A fair summary should provide a comprehensive coverage of diverse perspectives without underrepresenting certain groups. However, current work in summarization metrics and Large Language Models (LLMs) evaluation has not explored fair abstractive summarization. In this paper, we systematically investigate fair abstractive summarization for user-generated data. We first formally define fairness in abstractive summarization as not underrepresenting perspectives of any groups of people and propose four reference-free automatic metrics measuring the differences between target and source perspectives. We evaluate five LLMs, including three GPT models, Alpaca, and Claude, on six datasets collected from social media, online reviews, and recorded transcripts. Experiments show that both the model-generated and the human-written reference summaries suffer from low fairness. We conduct a comprehensive analysis of the common factors influencing fairness and propose three simple but effective methods to alleviate unfair summarization. Our dataset and code are available at https://github.com/psunlpgroup/FairSumm.
Recent advancements in Large Language Models empower them to follow freeform instructions, including imitating generic or specific demographic personas in conversations. We define generic personas to represent demographic groups, such as "an Asian person", whereas specific personas may take the form of specific popular Asian names like "Yumi". While the adoption of personas enriches user experiences by making dialogue systems more engaging and approachable, it also casts a shadow of potential risk by exacerbating social biases within model responses, thereby causing societal harm through interactions with users. In this paper, we systematically study "persona biases", which we define to be the sensitivity of dialogue models' harmful behaviors contingent upon the personas they adopt. We categorize persona biases into biases in harmful expression and harmful agreement, and establish a comprehensive evaluation framework to measure persona biases in five aspects: Offensiveness, Toxic Continuation, Regard, Stereotype Agreement, and Toxic Agreement. Additionally, we propose to investigate persona biases by experimenting with UNIVERSALPERSONA, a systematically constructed persona dataset encompassing various types of both generic and specific model personas. Through benchmarking on four different models -- including Blender, ChatGPT, Alpaca, and Vicuna -- our study uncovers significant persona biases in dialogue systems. Our findings also underscore the pressing need to revisit the use of personas in dialogue agents to ensure safe application.
Decisions made by machine learning models may have lasting impacts over time, making long-term fairness a crucial consideration. It has been shown that when ignoring the long-term effect, naively imposing fairness criterion in static settings can actually exacerbate bias over time. To explicitly address biases in sequential decision-making, recent works formulate long-term fairness notions in Markov Decision Process (MDP) framework. They define the long-term bias to be the sum of static bias over each time step. However, we demonstrate that naively summing up the step-wise bias can cause a false sense of fairness since it fails to consider the importance difference of different time steps during transition. In this work, we introduce a long-term fairness notion called Equal Long-term Benefit Rate (ELBERT), which explicitly considers varying temporal importance and adapts static fairness principles to the sequential setting. Moreover, we show that the policy gradient of Long-term Benefit Rate can be analytically reduced to standard policy gradient. This makes standard policy optimization methods applicable for reducing the bias, leading to our proposed bias mitigation method ELBERT-PO. Experiments on three sequential decision making environments show that ELBERT-PO significantly reduces bias and maintains high utility. Code is available at https://github.com/Yuancheng-Xu/ELBERT.