TrustLLM: Transparency

Abstract

Introduction

Background

Trust LLM Preliminaries 

Assessments

• Trustworthiness
• Truthfulness
• Safety
• Fairness
• Robustness
• Privacy Protection
• Machine Ethics 

Transparency

Accountability

Open Challenges

Future Work
Conclusions

--------------------------------------------------------------------------------------------------------

Original Content

Discussion of Transparency

Since LLMs can produce harmful content, spread misinformation, and have long-term environmental and socioeconomic consequences, transparency plays a central role in developing AI systems responsibly, ensuring that those involved can grasp what the model can and cannot do and how they operate and manage their outputs.

Responsible development and transparency go hand in hand in a world transformed by LLMs.

Some core transparency characteristics include balance opposite, increase in expectations, constant availability, and so on [622]. In this section, we begin by providing a summary of various perspectives in a broader context.

Subsequently, we delve into the specific dimensions of transparency concerning LLMs to explore the challenges they pose and the current research addressing these issues.

Different perspectives on transparency. It is worth noting that there is no universally accepted definition of transparency.

Transparency is a concept that has various dimensions, including information, normative, relational, and social perspectives [304, 623, 624]. In the following, we introduce transparency into three perspectives:

1) Informational transparency pertains to the disclosure of relevant details about a model or a system based on that model, ensuring a comprehensive understanding.

This emphasis on exposure aligns with the machine learning research community and industry best practices.

2) Normative transparency is a concept that regards transparency as a virtue and embodies a normative perspective by establishing criteria for assessing the conduct of public actors. [624]

3) In the context of relational and social transparency, transparency is not merely an attribute of an individual but rather a dynamic relationship between an agent and a recipient. It cannot be comprehended without this fundamental connection [625, 623].

This involves an institutional relationship facilitating the exchange of information concerning the workings or performance of an actor.

It is essential to acknowledge that these three perspectives are not entirely separate; they are interconnected but emphasize different aspects.

Related works. Research on improving the transparency of LLMs can primarily be categorized into two distinct approaches.

The first approach centers on increasing the transparency of the models themselves.

This is achieved through comprehensive documentation of both the models [626, 627] and the datasets [628, 629] upon which they are trained [304]. This method is practical and has gained widespread adoption in enhancing transparency for LLMs and other machine-learning models. Additionally, efforts have been made to advance transparency through designing and developing models with innovative architectures [630].

The second approach aims to enhance the transparency of the internal mechanisms and decision-making processes of LLMs. The Chain of thought paradigm [366] increases transparency by providing a detailed exposition of the intermediate steps and rationale employed by the model in formulating its conclusions.

This process significantly improves the interpretability of the model’s decision-making for human users [302]. Explainable AI [631] offers another pathway to transparency and explainability for LLMs by delivering frameworks and tools that demystify the internal circuits [632, 633], knowledge storing mechanisms [401, 402], and decision-making processes of these models [634].

Challenges. LLMs have evolved fast in recent years, developing unique attributes that set their transparency apart from other domains.

Many works have discussed the challenge to LLMs’ transparency. Overall, the challenge can be categorized into three main parts.

• Explainability of LLMs: A primary challenge hindering LLMs’ transparency is the underlying technology’s complexity. LLMs employ complex algorithms to predict the conditional probability of a token based on its contextual information, whether it’s a character, word, or another string. These contemporary LLMs rely on state-of-the-art neural network self-attention architectures like the transformer, boasting hundreds of billions or even trillions of parameters [635].
• In contrast to earlier models that operated on modest-sized datasets, LLMs are now trained on vast datasets containing hundreds of billions, or even trillions of tokens [393], necessitating significantly more computational resources and time. A fundamental pre-trained LLM serves as a versatile next-word predictor. Yet, LLMs offer the flexibility to be tailored to manifest or temper specific behaviors and enhance performance in distinct tasks such as text summarization, question answering, or code generation. This extensive scaling equips LLMs with significantly increased sophistication and expressiveness. However, this complexity also brings challenges when explaining their predictions.
• Participants adaptation: LLMs transparency often encompasses diverse participants, such as data scientists, model developers, executives, regulatory authorities, auditors, end-users, and individuals directly or indirectly impacted by a model or application [636]. Adopting LLMs may introduce fresh groups of participants with unique transparency concerns. However, it is crucial to recognize that transparency goes beyond simply sharing information; it also hinges on ensuring that the information is not only shared but comprehended and interpreted by the intended participants. Achieving genuine transparency through information disclosure requires adapting the information to cater to the specific needs of the participants [637].
• Public awareness: The evolving and often inaccurate public awareness of LLMs presents a challenge. Effective transparency strategies must account for the public’s existing cognitive framework, influenced by factors like mass media and language nuances. Addressing these flawed perceptions is crucial to prevent misuse and security risks, necessitating responsible information dissemination, in which organizations and the research community play a vital role in shaping public perception through their communication practices [638].

Diverse approaches, valuable insights.

There are a range of transparency-related approaches that have been investigated, by setting adaptive principles and mechanisms in different LLMs applying stages.

In the following, we provide a brief overview of these methods’ insights from different stages.

1)When architecting LLM applications, it is essential to consider the complexity of transparency from the beginning, including the transparency of the original pre-trained LLM, its adapted versions, and their integration into LLMinfused applications.

Maintaining clear distinctions between these components is imperative for achieving a comprehensive understanding of transparency within the realm of LLMs [639, 640].

Additionally, the LLM developers are responsible not only for providing information but also for considering the diverse participants who will receive and interpret that information [641].

2) When doing data processing, LLMs prompting, and fine-tuning, the developer needs to provide a clear explanation of the data being utilized, and the processing methods applied, and articulate the decision-making criteria, along with their justifications [642, 643].

3) Upon completing the utilization phase, developers should furnish a comprehensive model report, including information regarding model inputs and outputs, training methods, training data sources, developmental context, intended applications, and ethical considerations.

Furthermore, inspecting the system’s decisionmaking through audits should be enabled [627, 626].