Improving Prompt Consistency with Structured Generations
Hugging Face
Researchers at Hugging Face and Dottxt found that language model evaluation scores vary dramatically with minor prompt format changes—even though the information provided remains identical. Testing on MMLU showed individual models' performance swinging by 10 percentage points across different prompt formats, with Qwen1.5-7B dropping from 51.2% to 22.9% accuracy on the same task. Using structured generation to constrain model outputs reduced this variance and improved consistency in model rankings across different prompt variations.