[Author] Please place any models that need to be compatible with pdrop here.

[Cooperx521]

Thanks for attention!
I will address the compatibility issues of high-demand models as soon as possible after ICCV.

[zhangce01]

Hi, thank you for your great work! I am wondering if the proposed method would also be effective for the Qwen series models, such as Qwen2-VL and the more recent Qwen2.5-VL. I would greatly appreciate it if you could apply PDrop to these newer models and see how it works. Thanks!

[ashun989]

May I ask if you still have any plans to incorporate the PDrop application into an architecture such as Qwen2.5VL? @Cooperx521

[zhangce01]

May I ask if you still have any plans to incorporate the PDrop application into an architecture such as Qwen2.5VL? @Cooperx521

Hi @ashun989, over the past two months, we’ve implemented token pruning at the LLM decoding stage for Qwen-2.5-VL in our recent work, VScan (arXiv:2505.22654). The code is available at: https://github.com/Tencent/SelfEvolvingAgent/tree/main/VScan.

Our codebase also supports PyramidDrop through configurable hyperparameters, and can evaluate the effectiveness of various token pruning strategies on referring grounding tasks such as RefCOCO, RefCOCO+, and RefCOCOg with Qwen-2.5-VL. Feel free to reach out for further discussions!

[Cooperx521]

Sorry for the delayed response. I just finished some tasks recently, now starting to organize the PDrop code, and expect to support lmms-eval and Qwen2VL by next week. I'll let you know here once the code is released.

@ashun989 @zhangce01