Large language models (LLMs) like those behind chatbots and virtual assistants have become remarkably good at generating fluent and confident answers. But sometimes, these models provide responses even when they lack solid factual backing, which can lead to mistakes or misinformation. A newly published research paper explores a way to help LLMs decide when to answer a question and when it’s better to abstain, improving their reliability and trustworthiness.
Key Takeaways
- The study introduces “Chain-of-Self-Questioning” (CoSQ), a method that prompts LLMs to assess the information they need before committing to an answer.
- Using CoSQ, the models reduce their rate of incorrect answers by about 32% compared to a common reasoning approach called chain-of-thought prompting.
- CoSQ helps LLMs selectively abstain from answering when uncertain, leading to more accurate responses overall while still providing answers to most questions.
- The approach works consistently across eleven different language model families and various test conditions, demonstrating broad applicability.
The researchers focus on a key challenge in AI language models: these systems often generate confident-sounding answers even when they are unsure or lack evidence. To address this, they developed Chain-of-Self-Questioning (CoSQ), a “prompt-only” framework. This means it doesn’t require changing the model’s underlying architecture or retraining; instead, it guides the model through carefully designed prompts that encourage self-assessment.
CoSQ works by having the model internally ask itself a series of questions to identify exactly what information is necessary to answer the original query. Only if the model judges that it has enough support does it commit to an answer. Otherwise, it abstains, effectively saying, “I don’t know” or deferring the question. This selective answering is important because in many practical situations, making an unsupported claim can be more harmful than admitting uncertainty.
To test their approach, the team evaluated three variants of CoSQ on the TruthfulQA dataset, which consists of challenging multiple-choice questions designed to test models’ truthfulness. They ran experiments on eleven different language models, including both open-source and hosted options, under seventeen different settings. The best-performing CoSQ variant, called Grounded-CoSQ, cut the models’ wrong-commitment rate from 13.1% to 8.9%—a significant improvement—while boosting accuracy from 86.9% to nearly 90%. Importantly, it still answered 87.6% of questions, showing that abstaining did not come at the cost of excessive silence.
Other versions of CoSQ, named Critical-CoSQ and Adaptive-CoSQ, offered slightly different balances between how often the model answered and how reliably it did so. Across all variants and models, the approach consistently outperformed the usual chain-of-thought prompting baseline, which is a popular technique where the model explains its reasoning step-by-step before answering.
In addition to multiple-choice tests, the researchers ran a secondary evaluation on a dataset called Natural Questions Short-Answer, which involves open-ended responses. This helped confirm that the benefits of CoSQ extend beyond multiple-choice formats and into more natural, free-text answers.
By enabling language models to self-assess and selectively abstain, this research points toward safer and more trustworthy AI systems. In real-world applications like customer support, medical advice, or legal assistance, knowing when not to answer can prevent the spread of misinformation and reduce risks. The method’s prompt-only nature also means it can be applied without costly retraining, making it accessible for existing AI services.
Looking ahead, further work could explore how to integrate CoSQ with other AI safety techniques and how users perceive and respond to AI abstentions. As large language models become more widespread, tools like Chain-of-Self-Questioning may play a crucial role in balancing helpfulness with caution, ensuring these powerful systems remain reliable partners rather than sources of confusion.
Based on research published on arXiv by Ali Şenol.