We’re thrilled to share that Mohamed Dhouib, PhD candidate at ORAILIX, has had his paper "PACT: Pruning and Clustering-Based Token Reduction for Faster Visual Language Models" accepted at CVPR 2025. CVPR is one of the most prestigious conferences in computer vision!
This impactful research was conducted as part of the “Trustworthy and Responsible AI” chair, in collaboration with Crédit Agricole. PACT novel method that reduces the inference time and memory usage of Visual Language Models, contributing to the community's efforts to make multimodal LLMs usage more frugal. It relies on three steps: identifying and pruning unimportant visual tokens, merging redundant tokens using a new clustering algorithm (DBDPC), and retrieving key tokens to preserve information. DBDPC is a novel clustering algorithm that merges visually redundant tokens while keeping them within a defined distance threshold, preserving key visual information at low computational cost. After clustering, pruned tokens near cluster centers are reintegrated, and each cluster is merged into a single representative token to minimize information loss. Find out more on this topic via the following link: PACT: Pruning and Clustering-Based Token Reduction for Faster Visual Language Models Congratulations to Mohamed Dhouib on this major achievement, and thank you to our partners at Crédit Agricole for supporting this work through the Trustworthy and Responsible AI chair.