https://arxiv.org/abs/2009.03300

• 2021
• This article is already outdated. GPT-o1 is said to be excellent at STEM subjects (https://openai.com/index/learning-to-reason-with-llms/)

Key Contributions

1. Behavioral Testing Framework
   • Three Testing Types:
     • Minimum Functionality Tests (MFTs): Basic tests to ensure models handle essential cases (e.g., negation, antonyms).
     • Invariance Tests: Verify that outputs remain consistent when inputs undergo meaning-preserving changes (e.g., paraphrasing).
     • Directional Expectation Tests (DIRs): Confirm that outputs change in expected ways when input changes (e.g., increasing sentiment intensity).
2. Testing Matrix
   • CheckList evaluates across capabilities (e.g., vocabulary, negation handling) and domains (e.g., sentiment analysis, QA), ensuring systematic coverage.
3. User-Friendly Tooling
   • The framework provides an easy-to-use interface for crafting tests, automating them, and visualizing results.
4. Applications
   • CheckList was applied to widely used NLP models like BERT, RoBERTa, and GPT. It uncovered critical shortcomings, including failures in basic functionality (e.g., handling negations like "not happy") and domain-specific biases.

Key Findings

1. Models Lack Robustness
   • Despite high accuracy scores, state-of-the-art models struggle with simple linguistic variations and logical reasoning.
2. Behavioral Gaps
   • Many models fail to generalize beyond their training data, especially in edge cases or nuanced language scenarios.
3. Importance of Systematic Testing
   • Traditional evaluation metrics (accuracy, F1 scores) often give an incomplete picture of model performance. Behavioral testing helps uncover weaknesses that might otherwise go unnoticed.