2025-05-25-世界银行-平衡创新与严谨_人工智能评估的深思熟虑整合指南（指导说明）（英）_35页_1mb

Summary of "Balancing Innovation and Rigor: Guidance for the Thoughtful Integration of Artificial Intelligence for Evaluation"

Core Content

This guidance note explores the integration of large language models (LLMs) into evaluation practices, emphasizing the balance between innovation and analytical rigor. It outlines key considerations and good practices for effectively using LLMs in tasks such as structured literature reviews (SLRs) and evaluation synthesis.

Main Points

• LLMs in Evaluation: LLMs, a type of generative artificial intelligence, have the potential to improve the efficiency and scope of text data analysis in evaluation. They can perform well in tasks such as text classification, summarization, and synthesis.

• Challenges: Despite their strengths, LLMs may not always produce accurate or relevant outputs. This requires validation and refinement to ensure quality and alignment with evaluation objectives.

• Thoughtful Integration: The successful integration of LLMs into evaluation workflows depends on identifying relevant use cases, planning workflows, understanding resource allocation, and selecting appropriate metrics.

Key Considerations for Experimentation

Identifying Use Cases

• LLMs should be applied where they offer significant incremental value compared to traditional methods.
• Use cases must be aligned with the capabilities of LLMs and the specific needs of the evaluation.
• Examples of high-value use cases include text classification, summarization, and information extraction.

Identifying Opportunities Within Use Cases

• Complex tasks like SLRs can be broken down into smaller, manageable components.
• These components can be addressed using LLMs in a modular way, such as text search, manual review, classification, summarization, and synthesis.

Finding Agreement on Resources and Outcomes

• Teams must agree on the necessary resources (human, technological, and financial) and expected outcomes.
• This agreement helps manage expectations and ensures that the use of LLMs is not seen as a simple or inexpensive solution.

Selecting Appropriate Metrics to Measure LLMs' Performance

• For discriminative tasks (e.g., text classification), metrics such as recall, precision, and F1 scores are useful.
• For generative tasks (e.g., summarization, synthesis), subjective criteria like faithfulness, relevance, and coherence are more meaningful.
• Human judgments should be validated using metrics like Cohen's kappa to ensure intercoder reliability.

Experimental Results

• Text Classification: Achieved recall of 0.75 and precision of 0.60 on a test set of 30 papers. These scores were deemed satisfactory for the task.
• Text Summarization: Generated highly relevant (4.87), coherent (4.97), and faithful (0.90) abstracts, with no hallucinations observed.
• Text Synthesis: Produced a 500-word synthesis from six summaries, with all information correctly referenced (IC = 1.00, H = 1.00) and no hallucinations. However, relevance was slightly lower (4.20) due to some omissions.
• Information Extraction: Showed excellent faithfulness (IC = 1.00, H = 1.00) but had lower relevance (3.25), indicating difficulty in identifying the most relevant details.

Emerging Good Practices

• Representative Sampling: Divide the dataset into training, validation, testing, and prediction sets to ensure robustness and generalizability.
• Prompt Development: Start with a basic prompt and iteratively refine it based on model responses and feedback.
• Validation Loop: Implement an iterative process of prompt testing and refinement using validation and testing sets to ensure accuracy and reliability.
• Human Involvement: Manual review remains a mandatory component in workflows involving LLMs, ensuring that outputs are aligned with evaluation goals.

Conclusion

The guidance note highlights the importance of a structured, iterative, and rigorous approach when integrating LLMs into evaluation workflows. It provides a framework for identifying suitable use cases, developing effective prompts, and measuring model performance to ensure that LLMs are used responsibly and effectively.

Key Information

• LLMs can significantly enhance efficiency and validity in text-based evaluations.
• Validation is critical to ensure accuracy and alignment with evaluation objectives.
• Metrics vary by task type, with discriminative tasks using standard ML metrics and generative tasks using subjective criteria.
• Modular workflows allow for the reuse of successful components across different evaluations.
• Collaboration among multidisciplinary teams is essential for the responsible and effective use of LLMs in evaluations.