Large language models can still miss meaning hidden in plain sight, according to a new paper on arXiv. Even with significant advances in language understanding, these models fail to identify harmful intent in expressions that humans easily grasp. The authors describe this as a gap between human perception and model perception.

To address this, they propose a perception-aware adversarial self-play method. The idea is to have models generate and confront increasingly difficult examples, forcing them to recognize harmful content that would otherwise go unnoticed. This self-play loop is designed to expose and close the perception gap.

The work is presented as a step toward more robust harmful-content understanding. Because the paper is a single preprint, its claims have not yet been peer-reviewed, but the approach highlights a practical direction for improving safety in language systems.