Multi-object tracking in video sequences requires accurate object localization and consistent identity maintenance. While tracking-by-transformer methods have shown significant progress, current approaches employing a single decoder for both tracked and detection queries suffer from performance limitations due to task interference. Furthermore, the inherent discontinuity between an object’s previous and current states in a new frame contributes to association ambiguity. This work introduces a novel tracking-by-transformer framework incorporating explicit state transition prediction to mitigate these issues. By predicting object location and appearance changes between frames, our method reduces the state gap and improves association accuracy. Evaluation on the DanceTrack dataset, which features complex and rapid object motion, demonstrates substantial improvements: a 3.2% gain in AssA and a 3.0% gain in IDF1 compared to tracking-by-transformer model without state transition prediction.