LLMCheckup: Conversational Examination of Large Language Models via Interpretability Tools and Self-Explanations

Dialogue-based explanations are more effective than single explanations: The paper builds on prior research showing that conversational explanations help users better understand AI systems compared to one-off explanations, suggesting that interactive Q&A formats should be prioritized when designing AI transparency tools.

LLMs can explain their own behavior without additional training: LLMCheckup demonstrates that large language models can generate meaningful explanations and recognize user questions about their reasoning through self-explanations and rationale generation, without requiring fine-tuning or specialized external modules—making interpretability more accessible.

Multiple explanation methods work better together: The tool combines different explainability approaches (feature attributions, self-explanations, and rationale generation) into one interface, suggesting that responsible AI oversight benefits from layered explanation methods rather than relying on a single technique.

User expertise varies, so tools must be adaptable: The system includes tutorials for users with different levels of XAI knowledge and supports multiple input types, highlighting that responsible AI tools need to serve both technical experts and non-specialists to be truly useful for accountability.

Practical applications demonstrate real-world value: The paper demonstrates LLMCheckup on fact-checking and commonsense reasoning tasks, showing that conversational AI examination tools can be applied to concrete, high-stakes use cases where understanding model behavior matters.