A prompt prefix makes AI multiple-choice answers more reliable — and cheaper than full reasoning

Researchers found that prepending a simple phrase like 'The correct option is:' to an AI model's output makes it pick the right multiple-choice answer far more often, without retraining the model. This method is faster and cheaper than asking the AI to reason through the problem fully, and works across different models and question sets.

AI systems are increasingly evaluated on multiple-choice tests by checking which answer the model thinks is most likely on its first generated token — a fast, cheap method that doesn't require full reasoning. But it's fragile: models often assign high probability to random tokens or use valid answer letters as throwaway filler rather than actual answers, poisoning the evaluation. This paper shows a trivial fix that actually works: a structured prefix steers the model to give clean answers without any model modification. The implication is immediate and practical: if you're building systems that need to answer multiple-choice questions at scale (hiring tests, compliance certification, educational software), you can now use the cheap evaluation method and get reliability close to the expensive reasoning method. That changes the cost-performance tradeoff for large-scale deployment.