Is learnable_query updated?

[zzl-cv]

This is impressive work that has provided me with much inspiration. I noticed that during both the first stage (3b_stage1.sh) and the second stage (3b_cot.sh) of training, all other parameters are being updated, but the saved weights for learnable_query remain unchanged. This suggests that the learnable_query module may not be undergoing training.

Has the author encountered this phenomenon before? I would greatly appreciate any insights or explanations.

[catchcodes]

In our experiments, the learnable_query is indeed updated normally during training, and we did not observe cases where its weights remain unchanged. We verified this by directly comparing learnable_query parameters across checkpoints saved at different iterations.

Although the element-wise differences may appear close to zero at first glance (especially when using bfloat16 precision), quantitative comparisons such as the maximum difference and the L1 difference clearly show non-zero updates, as illustrated in the figure above.