One Token per Multimodal Evidence: Latent Memory for Resource-Constrained QA

In three linesLatent Memory replaces each memory item (text/image) with a single compressed latent token, reducing generator token consumption by 3-10x. Trained with reconstruction, contrastive, and distillation objectives, the system achieves competitive performance on HotpotQA and multimodal benchmarks while lowering memory pressure.

Summary generated by Claude — human-verified