LatentPort: Beyond KV Cache - Cross-Model Transfer of Recurrent Memory in Hybrid Language Models: A 4B-to-9B Hybrid-State Handoff Without Target Prefix Replay
First demonstrated cross-model handoff of persistent recurrent inference state between Qwen3.5 4B and 9B hybrid models without target prefix replay.
LatentPort demonstrates persistent hybrid-state transfer from a Qwen3.5 4B model to a 9B sibling without the receiver rereading context, claimed as the first cross-model handoff of recurrent inference state without target prefix replay. Adding the Gated DeltaNet persistent-state package to translated attention KV lowers teacher-forced NLL by 0.747 nats/token, improving all 64 PG19 documents. A 434,176-parameter correction yields 0.076 nats/token excess NLL versus native 9B, JS divergence 0.022, and native context recovery of 0.918, though evidence covers only one direction, one geometry-matched pair, and 4K teacher-forced continuation.
- First 4B-to-9B recurrent-state handoff with zero target prefix tokens processed
- GDN persistent-state package cuts NLL 0.747 nats/token across 64 PG19 documents
- Corrected 9B reaches 0.076 nats/token excess NLL, 0.918 native context recovery
- Scope limited: one direction, one model pair, 4K continuation, 16K untested
Full article204 words · extracted from huggingface.co · click to collapse
Can one language model hand its live memory to another without the receiver rereading the context? We demonstrate useful persistent hybrid-state transfer across one architecture-matched Qwen3.5 4B-to-9B sibling pair. To our knowledge, this is the first demonstrated cross-model handoff of persistent recurrent inference state between differently sized hybrid language models without target prefix replay. Translated attention KV alone leaves a large gap; adding the Gated DeltaNet (GDN) persistent-state package lowers teacher-forced negative log-likelihood (NLL), the average next-token log-loss, by 0.747 nats/token (95% paired document bootstrap CI [0.6921, 0.8047]), improving all 64 PG19 documents. Direct recurrent and convolution reuse outperforms the tested learned GDN maps, consistent with partial functional compatibility of persistent-state coordinates. A fresh component factorial selects translated KV with direct recurrent and convolution state. An additional 434,176-parameter correction improves that base on 64 fresh web documents: continuation loss is 0.076 nats/token above native 9B (excess NLL), Jensen-Shannon (JS) divergence is 0.022, and native context recovery (NCR) is 0.918. Corrected 9B significantly beats continued 4B inference while processing zero historical prefix tokens. Evidence 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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.25053