pope-benchmark

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.

A controlled evaluation audit of the POPE benchmark. We prove that prompt templates and answer-extraction rules cause an 86-point F1 swing (0.000 to 0.868) even when model outputs are completely frozen.