Investigating Data Leakage–Induced Over-Confidence and Explanation Faithfulness in Transformer-Based Text and Audio Models

Transformer-based models now sit at the center of modern artificial intelligence, powering systems that read, listen, respond, and increasingly act on behalf of humans. From large-scale language models to voice-driven assistants, these systems exhibit a level of fluency and responsiveness that would have seemed implausible only a few years ago. Their success has been driven largely by scale Larger models, larger datasets, and longer training regimes resulting in impressive performance across text and audio tasks.