OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always!

Researchers introduced OffTopicEval, a benchmark revealing that all major LLMs suffer from poor operational safety, with even top performers like Qwen-3 and Mistral achieving only 77-80% accuracy in staying on-topic for specific use cases. The study proposes prompt-based steering methods that can improve performance by up to 41%, highlighting critical safety gaps in current AI deployment.

• →All evaluated LLMs show significant operational safety failures, with even the best models achieving only 77-80% accuracy in appropriate query handling.
• →GPT models plateau at 62-73% operational safety scores, while Llama-3 performs poorly at just 23.84%.
• →Prompt-based steering methods like Q-ground and P-ground can substantially improve safety, with gains up to 41%.
• →Operational safety represents a fundamental challenge for enterprise LLM deployment beyond generic harm considerations.
• →The research highlights urgent need for safety interventions before wide-scale LLM agent deployment.