Reproduction Results: AICG Module Shows No Improvement Over Direct Attention Fusion

[zyf-815]

I have reproduced your work and observed that the AICG module does not produce the expected performance gains. In my experiments, directly applying Reference Attention without the AICG mechanism achieves better results than the proposed AICG-based approach.

[hangfrieddays]

Thanks for reproducing our work and for sharing the result.
From our original recorded runs, we did not observe this behavior. However, the training pipeline itself contains multiple stochastic components, including random cropping, online degradation, and shuffling, so some variation across different random seeds is expected.