Consistent real-world interpretability of LLMs under diverse environmental conditions

Establish methodologies to achieve consistent real-world interpretability of large language models across diverse environmental conditions in autonomous driving applications, ensuring that model decisions and explanations remain understandable and reliable when weather, visibility, and related environmental factors vary.

Background

The paper evaluates LLMs on AgentDrive-MCQ, including a scenario-style reasoning category that requires holistic understanding of complex, dynamic environments. Frontier models achieved near-perfect scenario reasoning scores, indicating strong situational awareness and context synthesis.

Despite these strong results, the authors explicitly state that achieving consistent real-world interpretability across diverse environmental conditions remains an open research challenge, highlighting a critical gap between benchmark performance and reliable interpretability in varied operational settings.