TwigVLM introduces a lightweight “twig” module that prunes visual tokens early and enables self‑speculative decoding, achieving up to 154% speedup on long‑text generation while preserving 96% of original LVLM accuracy, as demonstrated on LLaVA‑1.5‑7B and other benchmarks.
This article presents Dynamic Vision Mamba (DyVM), a method that tackles token and block redundancy in Mamba‑based visual models through a novel re‑ordering pruning strategy and dynamic block selection, achieving a 35.2% FLOPs reduction with only a 1.7% accuracy loss while demonstrating strong generalization across tasks and architectures.
This paper investigates the “hourglass” phenomenon in residual‑quantized semantic identifiers for generative search and recommendation, revealing that token concentration in intermediate codebooks causes path sparsity and long‑tail distributions, and proposes heuristic layer removal and adaptive token‑pruning strategies that markedly improve model performance.
