UAV-ESPrune: Evidence-Aware Visual Token Pruning for UAV Vision-Language Reasoning

Low-altitude unmanned aerial vehicles (UAVs) require compact multimodal inference for perception, environment understanding, and planning-oriented reasoning. Vision-language models (VLMs) provide a shared interface for these tasks, but long visual-token sequences increase language-model prefill computation, sequence-dependent memory traffic, and visual key-value (KV) storage. Generic pruning often treats feature similarity as redundancy. In UAV imagery, similar patches may represent independent small-object instances, structural transitions may affect planning decisions, and distributed scene context may support environment classification. UAV-ESPrune is a training-free, before-LLM framework that reduces visual tokens while preserving such evidence. It combines visual saliency, local peak evidence, structural-transition cues, peak-budget anti-collapse, weighted geometric-feature coreset selection, and residual condensation. Experiments use Qwen2.5-VL-7Bon three selected MM-UAVBench subtasks: referring expression counting, environment state classification, and ground target planning, with token retention from 15% to 75%. At 25% retention, UAV-ESPrune retains the full-token counting accuracy of 30.86%, reaches 74.92% for environment classification, and obtains 40.62% for planning. At this retention ratio, independently measured end-to-end inference time decreases from 671.67 to 591.47 ms, a reduction of 11.94%. Visual KV storage decreases by 75.2% relative to full-token inference. These results support evidence-preserving token reduction for UAV vision-language reasoning.