#attention-weighted

Clustering and Token Denoising for Faster and More Robust VLMs⭐8

ClustRS reduces VLM tokens by 97% while improving noise robustness. This training-free method boosts accuracy by up to 20% on LLaVA benchmarks.