0:00
/
0:00
Transcript

"Multi-expert Prompting Improves Reliability, Safety, and Usefulness of Large Language Models"

The podcast on this paper is generated with Google's Illuminate.

Multiple AI experts working together make better decisions than a single expert

https://arxiv.org/abs/2411.00492

🤖 Original Problem:

Single-expert prompting frameworks like ExpertPrompting can introduce bias and limit perspectives when handling open-ended queries. They often provide one-sided views, missing the depth needed for comprehensive answers.

-----

🛠️ Solution in this Paper:

→ Introduces Multi-expert Prompting with two main steps:

The method has two main steps:

1. Expert & Response Generation - Generates n expert identities with concise role descriptions and gets responses from each expert

2. Expert Response Aggregation - Uses 7 carefully designed subtasks based on Nominal Group Technique to combine expert responses and select the best output.

→ Uses zero-shot prompting to generate diverse expert responses in parallel

→ Implements a novel 7-subtask method to:

- Identify agreed viewpoints

- Handle conflicting opinions

- Capture unique perspectives

- Select best response through systematic evaluation

-----

💡 Key Insights:

→ Three experts yield optimal results for truthfulness and factuality

→ Aggregating expert responses in a single turn is more efficient than iterative refinement

→ Human-designed NGT framework outperforms LLM-generated plans for response aggregation

→ Diverse domain experts working in parallel produce better results than sequential expert consultation

-----

📊 Results:

→ Improves truthfulness by 8.69% with ChatGPT over best baseline

→ Achieves state-of-the-art truthfulness scores on TruthfulQA-Generation

→ Wins 75% cases for informativeness and 76.5% for usefulness

→ Completely eliminates toxic content and reduces hurtfulness

Discussion about this video