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
A smaller Qwen hybrid model passed usable live recurrent state to a larger sibling without making it reread the prompt.
The paper reports a 4B-to-9B handoff in one architecture-matched Qwen3.5 pair, with zero target prefix replay. KV transfer alone left a large loss gap, while adding persistent Gated DeltaNet state improved next-token loss across all 64 PG19 test documents. A small correction module brought the 9B continuation close to native 9B behavior on fresh web documents and beat continued 4B inference. The author limits the claim to this one matched model pair and direction; longer-context results, free-generation equivalence, and a general state interface remain unproven. HF Daily Papers' note
The paper reports a 4B-to-9B handoff in one architecture-matched Qwen3.5 pair, with zero target prefix replay. KV transfer alone left a large loss gap, while adding persistent Gated DeltaNet state improved next-token loss across all 64 PG19 test documents. A small correction module brought the 9B continuation close to native 9B behavior on fresh web documents and beat continued 4B inference. The author limits the claim to this one matched model pair and direction; longer-context results, free-generation equivalence, and a general state interface remain unproven. HF Daily Papers' note
score 5