Chain-of-Self-Questioning (CoSQ): Enabling LLMs to Abstain Responsibly
Large Language Models (LLMs) can sometimes confidently generate incorrect information, especially when their internal knowledge is weak. Chain-of-Self-Questioning (CoSQ) is a novel framework that allows LLMs to assess their own knowledge and decide whether to answer a question or abstain, thereby improving reliability and reducing the risk of propagating misinformation.
Reading in Plain English mode — toggle below to switch
Large Language Models (LLMs) can sometimes confidently generate incorrect information, especially when their internal knowledge is weak. Chain-of-Self-Questioning (CoSQ) is a novel framework that allows LLMs to assess their own knowledge and decide whether to answer a question or abstain, thereby improving reliability and reducing the risk of propagating misinformation.
Imagine a student who, before answering a difficult exam question, first asks themselves: 'Do I have enough information to answer this confidently?' If not, they might choose to say 'I don't know' rather than guessing incorrectly. CoSQ gives LLMs a similar self-assessment capability.
How CoSQ Works
- 11. **Initial Question Analysis**: The LLM first processes the input question to understand its core requirements.
- 22. **Self-Question Generation**: Based on the analysis, the LLM generates a set of internal 'self-questions' to probe its own knowledge base regarding the information needed to answer the original question accurately.
- 33. **Self-Answer Generation**: The LLM then attempts to answer these self-questions, effectively checking its internal state and factual support.
- 44. **Confidence Assessment**: Based on the quality and completeness of its self-answers, the LLM assesses its confidence in providing a correct answer to the original question.
- 55. **Decision to Answer or Abstain**: If the confidence score meets a predefined threshold (τ), the LLM provides an answer. Otherwise, it abstains, indicating it cannot confidently answer.
P(Answer | Confidence ≥ τ)
The probability of an LLM providing an answer is conditional on its self-assessed confidence being greater than or equal to a predefined threshold (τ).
In medical diagnosis, an LLM using CoSQ might be asked about a rare disease. If its self-assessment reveals insufficient or conflicting information, it would abstain from providing a definitive diagnosis, preventing potentially harmful misinformation and prompting a human expert consultation.
CoSQ significantly reduces the rate of wrong commitments by LLMs while maintaining or improving accuracy, allowing for tunable control over risk in critical applications.