Megadose The AI curve, compressed.

Out of Sight, Into Memory: Compact Visual Latents Retain Local Evidence in Ultra-High-Resolution Remote Sensing

· ArXiv · AI/CL/LG ·
A 64-token visual memory recovered most of the accuracy gain from detailed image crops while using 7.6–9.9 times fewer retained tokens.

Across two generations of Qwen-VL on LRS-GRO, accuracy rose by 2.5–4.8 percentage points over a downsampled global image alone. Results on XLRS-Bench showed consistent behavior. Swapping in memory from another image or region cut accuracy by 5.8–7.4 points, performing worse than removing memory altogether. The tests isolate evidence retention from region selection: they show that compressed memory preserves useful local detail, without testing how models choose where to look.

ArXiv · AI/CL/LG's note

score 4

Categories: Research