AdaptInfer: Adaptive Token Pruning for Vision-Language Model Inference via Dynamical Text Guidance

Name: AdaptInfer: Adaptive Token Pruning for Vision-Language Model Inference via Dynamical Text Guidance
Uploaded: 2026-10-10T04:17:43.786106+00:00
Duration: 342 s
Description: AdaptInfer is a plug-and-play framework for pruning vision tokens in Vision-Language Models (VLMs) during inference. It uses dynamic text guidance, leveraging layer-wise text attention maps to estimate token importance and create an efficient pruning schedule based on cross-modal attention shifts, significantly reducing latency while maintaining high accuracy.

AdaptInfer is a plug-and-play framework for pruning vision tokens in Vision-Language Models (VLMs) during inference. It uses dynamic text guidance, leveraging layer-wise text attention maps to estimate token importance and create an efficient pruning schedule based on cross-modal attention shifts, significantly reducing latency while maintaining high accuracy.

Open the full episode