There are some very interesting latent compaction approaches like this[1] for when you can control the whole inference stack. i.e in on-device and datacenter in

There are some very interesting latent compaction approaches like this[1] for when you can control the whole inference stack. i.e in on-device and datacenter inference. [1]: https://arxiv.org/abs/2602.16284 https://arxiv.org/abs/2602.16284