This research introduces a framework for making language models consistently perform well even when users phrase the same question in different ways.

Key Findings:

• LLMs show significant performance degradation when faced with semantically equivalent but differently phrased prompts
• The proposed framework focuses on worst-case performance across paraphrases rather than average performance
• Improves prompt robustness without computationally expensive inference-time algorithms
• Provides a systematic approach versus traditional trial-and-error prompt engineering

Security Implications: By enhancing robustness against paraphrased inputs, this research directly addresses reliability concerns in LLM deployments, reducing vulnerability to adversarial manipulations and ensuring more consistent security profiles across varied but semantically equivalent user inputs.

Same Question, Different Words: A Latent Adversarial Framework for Prompt Robustness