multimodalfactbullishA single learnable embedding trained on small image-QA datasets can effectively select query-relevant visual tokens from pre-filled KV caches to handle long visual contextsMachine Learning02 Aug 2026http://arxiv.org/abs/2607.28627v1