LatentPort, a new paper posted on Hugging Face, asks whether one language model can hand its live memory to another without the receiver rereading the context. The authors demonstrate what they describe as the first cross-model handoff of persistent recurrent inference state between differently sized hybrid language models without target prefix replay.
The experiment uses an architecture-matched Qwen3.5 4B-to-9B sibling pair. The 4B model processes a 4,096-token context and then transfers its inference state — including translated attention KV and the model's persistent recurrent hidden state — to the 9B model, which receives zero historical context tokens.
According to the paper, translated attention KV alone leaves a large gap. Adding the Gated DeltaNet (GDN) persistent-state package lowers teacher-forced negative log-likelihood (NLL) by 0.747 nats/token, with a 95% paired document bootstrap confidence interval of [0.6921, 0.8047], improving all 64 PG19 documents. Direct recurrent and convolution reuse outperformed the tested learned GDN maps, which the authors read as evidence of partial functional compatibility of persistent-state coordinates.
A fresh component factorial selected translated KV with direct recurrent and convolution state. An additional 434,176-parameter correction improved that base on 64 fresh web documents: continuation loss came in at 0.076 nats/token above native 9B (excess NLL), Jensen-Shannon divergence was 0.022, and native context recovery (NCR) reached 0.918. The corrected 9B significantly beat continued 4B inference while processing zero historical prefix tokens.
The authors are explicit about the limits of the evidence: it covers one direction, one geometry-matched Base-model pair, and 4K teacher-forced continuation. The near-native gate failed, the 16K branch was not run, and free-generation equivalence and a general state interface remain unproven. The paper frames the broader question as whether models could eventually expose a state interface that lets computation and memory move directly between models instead of being converted back into tokens.