It's actually a common, if poorly discussed, phenomenon in research: a careful reimplementation of a model often outperforms the original paper's results. This usually isn't due to a secret breakthrough, but to the quiet, unglamorous work of fixing hidden bugs and standardizing evaluation.

The reasons are almost always practical. Small oversights in the original code, like misaligned video frames, inconsistent data preprocessing, or even a handful of missing samples, can subtly hamper performance. When these are corrected in a reimplementation, the model finally performs as it was always meant to, making it seem "better."

However the biggest factor is the evaluation methodology itself. Seemingly minor choices, like how to average F1 scores across videos or which random seed to use, can swing results by several percentage points. A higher score might not mean a smarter model, but a more precise and consistent measurement.

So, an improved score should be seen less as a new claim and more as a sign of the field's reproducibility challenges. The most valuable contribution in such cases is transparency: meticulously documenting what was changed, clarifying the evaluation protocol, and sharing the code. This allows the community to understand whether the gain is a genuine step forward or simply the result of a more rigorous and careful setup.

Reimplementation may yield higher F1 scores than the reported baseline due to differences in preprocessing steps, hyperparameter tuning, updated libraries, or computational environments. Improved optimization methods and hardware can also enhance training stability and performance. Additionally, subtle variations in random dataset splits may lead to better generalization, making the reimplemented model outperform the original baseline.