How much LLMs get confused by different ways of asking the same question.
Meet ProSA: The framework that catches LLMs playing favorites with prompts
Original Problem 🔍:
LLMs exhibit prompt sensitivity, affecting performance and user satisfaction. Current research overlooks instance-level variations and subjective evaluations.
Solution in this Paper 🛠️:
• ProSA framework for evaluating prompt sensitivity in LLMs
• Novel PromptSensiScore (PSS) metric for quantifying sensitivity
• Instance-level analysis across multiple tasks and datasets
• Utilizes decoding confidence to explain underlying mechanisms
Key Insights from this Paper 💡:
• Prompt sensitivity varies across datasets and models
• Larger models generally show enhanced robustness
• Few-shot examples alleviate prompt sensitivity
• Subjective evaluations are susceptible to prompt sensitivities
• Higher model confidence correlates with increased prompt robustness
Results 📊:
• Llama3-70B-Instruct demonstrates highest robustness
• Transition from zero-shot to one-shot shows significant improvement
• Larger LLMs benefit more from increased few-shot instances
• LLMs robust in answering straightforward queries, sensitive in complex tasks
• Prompt sensitivity reflects model's confidence level
🧠 The analysis revealed that prompt sensitivity is essentially a reflection of the model's confidence level:
Higher confidence in outputs correlates with increased robustness against prompt semantic variations
When a model is robust to prompts for a given instance (low PSS score), it exhibits the highest decoding confidence
Conversely, when sensitive to prompts, the model's decoding confidence decreases