mlsanityData to 7 Oct 2026

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

arXiv 2609.25053 ↗First posted 7 Sep 2026cs.CL cs.AIcited by 0

On the map: Efficient LLM inference and serving → KV cache optimization

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in KV cache optimization, one of 12 directions in Efficient LLM inference and serving. Builds on 19 of its 24 references that are in the map. First posted on arXiv 7 Sep 2026.

Builds on

19 of 24 references.

Closest ancestors · 5

Same direction, newest first

  1. 2026CacheBridge: Efficient Cross-Model KV Cache Transfer1 Sep
  2. 2026Tail-Replay: Escaping the Curse of Linear Attention in Prefix Caching for Hybrid LLMs31 Aug
  3. 2026Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse4 Aug
  4. 2026KV Cache Translation across Heterogeneous Large Language Models31 Jul
  5. 2025Cache-to-Cache: Direct Semantic Communication Between Large Language Models3 Oct

Other references · 14

  1. 2024The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scalecited by 1,104
  2. 2019Compressive Transformers for Long-Range Sequence Modellingcited by 699
  3. 2024Gated Delta Networks: Improving Mamba2 with Delta Rulecited by 483
  4. 2024Marconi: Prefix Caching for the Era of Hybrid LLMscited by 29
  5. 2026Latent Cache Flow: Model-to-Model Communication Without Textcited by 6
  6. 2026StateBridge: Training-free Hidden-state Alignment for Latent Communication in LLM Multi-Agent Systemscited by 5
  7. 2026Semantic Cache Distillation: Efficient State Transfer via Reuse and Selective Patchingcited by 4
  8. 2026Architecture-Dependent Causal Transfer of Activation States Across Large Language Modelscited by 3
  9. 2026Frozen Memory Is Not Enough: Rethinking External Memory as Extractioncited by 3
  10. 2026S0 Tuning: Zero-Overhead Adaptation of Hybrid Recurrent-Attention Modelscited by 2
  11. 2026DASC: Decay-Aware State Compression for Hybrid Linear-Attention Servingcited by 2
  12. 2026WriteSAE: Sparse Autoencoders for Recurrent Statecited by 1
2 more
  1. 2026Dual-Cache Latent Space Communication between Heterogeneous Language Modelscited by 1
  2. 2026A Universal Context-Reuse Layer for Cross-Model KV Sharingcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026What Attention Recalls and Recurrence Controls in Hybrid Language Models
  2. 2026Memory Canonicalization: A Framework and Benchmark for Cross-Model Drift in Persistent LLM Memory
  3. 2026HeadWiseKV: Budgeted Per-Head Cache Residency for Hybrid Long-Context Language Models
  4. 2026DAMP: Decay-Aware Mixed-Precision Recurrent-State Quantization
  5. 2025Stateful KV Cache Management for LLMs: Balancing Space, Time, Accuracy, and Positional Fidelity
  6. 2026Balancing Memory Pathways: Analyzing and Improving Memory Utilization in Hybrid LMs
  7. 2026DreamingGoose: Staged Distillation from Autoregressive Transformers to Bidirectional Recurrent Diffusion Language Models
  8. 2026Expressivity-Efficiency Tradeoffs for Hybrid Sequence Models