Query-aware visual token pruning for VLMs. Five-component pipeline (cross-attention scorer, entropy controller, spatial coherence, progressive schedule, token recycling) attached via PyTorch forward hooks — no upstream model changes. Evaluated on POPE with LLaVA-1.5-7B.
pytorch inference-acceleration kv-cache token-pruning llava training-free vision-language-models- pope-benchmark
-
Updated
May 11, 2026 - Python