Appier Study: Teaching AI to Recognize Insufficient Information and Choose the Right Reasoning Language

Appier’s AI research team published two papers. After testing 28 leading LLMs, they found that when the correct answer was “none of the above,” model accuracy fell by 30% to 50%. After DPO training, models’ accuracy in identifying questions with no correct answer improved by nearly 30 percentage points. Another study found that large reasoning models often default to reasoning in high-resource languages such as English. For some models, the reasoning language differed from the response language in more than 90% of cases. Reasoning in local languages was more effective for cultural understanding and safety detection.