Model framework. It employs a multi-round competitive selection mechanism. The process begins with input tokens (gray squares). In long sequences, the standard softmax function introduces a competition that can diminish the informational contribution of many tokens to near zero. To counteract this and decouple the selection process (“competition”) from the intrinsic value of the information, our model iteratively identifies and discards less salient features through multiple rounds. The “winner” tokens from each round are highlighted within the boxes outlined with blue dashes. The final set of “Winner” tokens, representing the most critical local information, is then integrated with the initial sequence to enrich the model’s overall perceptual field.