Multi-expert Prompting Improves Reliability, Safety, and Usefulness of Large Language Models (2411.00492v1)

Abstract: We present Multi-expert Prompting, a novel enhancement of ExpertPrompting (Xu et al., 2023), designed to improve the LLM generation. Specifically, it guides an LLM to fulfill an input instruction by simulating multiple experts, aggregating their responses, and selecting the best among individual and aggregated responses. This process is performed in a single chain of thoughts through our seven carefully designed subtasks derived from the Nominal Group Technique (Ven and Delbecq, 1974), a well-established decision-making framework. Our evaluations demonstrate that Multi-expert Prompting significantly outperforms ExpertPrompting and comparable baselines in enhancing the truthfulness, factuality, informativeness, and usefulness of responses while reducing toxicity and hurtfulness. It further achieves state-of-the-art truthfulness by outperforming the best baseline by 8.69% with ChatGPT. Multi-expert Prompting is efficient, explainable, and highly adaptable to diverse scenarios, eliminating the need for manual prompt construction.

• The paper demonstrates that multi-expert prompting significantly improves LLM reliability and safety by aggregating responses from simulated experts.
• It employs a two-step methodology—expert response generation and aggregation—to reconcile diverse viewpoints effectively.
• Evaluation on benchmarks shows up to an 8.69% improvement in truthfulness and a notable reduction in toxic content over standard models.

Overview of Multi-expert Prompting for LLMs

The paper "Multi-expert Prompting Improves Reliability, Safety and Usefulness of LLMs" by Do Xuan Long et al. presents a strategy aimed at enhancing the reliability and overall performance of LLMs. This method, termed Multi-expert Prompting, seeks to enrich the generation quality of LLMs by simulating and aggregating responses from multiple experts. The technique draws from established decision-making paradigms such as the Nominal Group Technique (NGT), and it has been shown to outperform existing baselines by significant margins in key performance areas.

Core Methodology

The primary aim of Multi-expert Prompting is to guide LLMs to produce more reliable and multifaceted outputs by leveraging simulated expert roles. The process is divided into two pivotal steps:

1. Experts Response Generation: The model is prompted to generate diverse expert identities along with succinct descriptions pertinent to a given task. These simulations are informed by zero-shot prompting, negating the need for complex demonstrations, and allowing the LLM to provide independent responses via these expert personas.
2. Expert Response Aggregation: This step synthesizes and evaluates these individual responses through a series of well-defined subtasks, echoing the NGT framework. These include consensus-building among similar responses, resolution of conflicting statements, and the fusion of unique viewpoints. The most complete, factual, and useful response is selected as the final output after aggregation.

Evaluation and Results

Multi-expert Prompting has been rigorously evaluated against several benchmarks such as TruthfulQA, FactualityPrompt, BOLD, and HONEST to test truthfulness, factuality, and safety aspects. The results demonstrate the proposed method's superiority in producing truthful (with an improvement of up to 8.69%) and non-toxic content, marking a considerable advancement over single-expert frameworks. Furthermore, the paper provides a comprehensive assessment of the method's informativeness and usefulness in open-ended scenarios, where it consistently outstrips existing baselines.

Critical Insights and Implications

A significant contribution of this methodology lies in its ability to systematically integrate multiple viewpoints effectively, which is critical in addressing biases and misinformation that can often skew LLM outputs. This incorporation of varied perspectives reduces the systemic failure modes observed in earlier single-expert modeling practices.

In practice, Multi-expert Prompting could have lasting impacts on the deployment of LLMs in domains requiring nuanced decision-making and enriched information reporting, such as healthcare, policy-making, and education. The theoretical grounding of this approach underscores the value of structured expert input, fostering a more balanced and less biased reinforcement mechanism in AI systems.

Future Directions

The authors suggest several pathways for future research. These include exploring further generalizations of expert role configurations, optimizing aggregation subtasks for complex reasoning, and incorporating adaptive weighing mechanisms that can prioritize more contextually relevant expert opinions. The application scope could also extend to broader AI-driven group decision-making tasks, potentially transforming how automated insights are utilized in professional environments.

Overall, Multi-expert Prompting introduces a promising paradigm shift in the generation capabilities and alignment of LLMs with human intent, emphasizing reliability, safety, and a deeper engagement with diverse knowledge structures. This stands as a testament to the ongoing evolution of AI methodologies in producing more accountable and informed machine intelligence responses.