Transformer architecture that dynamically halts tokens at inference

This invention describes a way for AI models, specifically those called "transformers" used for finding objects in sensor data (like from a car's cameras or radar), to work more efficiently. It does this by letting the model decide that some pieces of information (called "tokens") are not important enough to process further, effectively "halting" them. The model then focuses its computational power only on the important pieces to detect objects faster, specifically by identifying tokens with a low score.