Embodied AI agents increasingly need to run several tasks at once—manipulating objects, conversing, and building memory—while sharing the same observations. These tasks often have different time constraints, which complicates inference for vision-language-action (VLA) models. A new arXiv preprint, OxyGen, takes on this problem with a unified approach to KV cache management.

The paper (arXiv:2603.14371v4) is aimed at Mixture-of-Transformers architectures, a recent direction for VLA models. Rather than treating each task's inference separately, OxyGen manages the KV cache—the memory transformers use during generation—across parallel tasks. The abstract describes supporting manipulation, conversation, and memory construction from shared observations under distinct time constraints.

This is a single preprint, and the abstract does not give implementation details or benchmark results. Still, the focus on multi-task parallelism for embodied AI points to a practical bottleneck: making a single model serve many real-time responsibilities at once.