The paper introduces Hybrid State Transfer (HST), a method for transferring not only full‑attention key/value caches but also recurrent (REC) and convolution (Conv) states between different large language models during long conversations. HST preserves the order of Conv histories and uses structured REC maps, calibrating adapters with the target model’s outputs to maintain continuity. Experiments on five benchmarks show that Qwen3.5 models achieve high task retention (up to 94%) and achieve over 21× speedups on synthetic 16K inputs when handoff is performed.
arXiv:2610.15509v1 Announce Type: cross Abstract: KV caches make long-context serving efficient by reusing past computation, but model switching breaks this reuse and forces the receiver to prefill t...
arXiv:2608.30963v1 Announce Type: cross Abstract: Modern large language model (LLM) serving systems increasingly operate over repeated or shared context, yet each model typically performs its own pre...
The paper introduces an LSTM-style transformer that shares a single key–value cache across its recurrent core, reducing parameters by 40.4% compared to an untied model while keeping forward FLOPs similar. It aligns parallel and continuation training passes with a consistency loss, eliminating 81–89% of the continuation‑loss penalty on FineWeb. Experiments across multiple widths and token budgets show that weight tying incurs the smallest loss cost at lower token budgets, highlighting its potential for data‑ and storage‑constrained settings.
arXiv:2608. 11231v1 Announce Type: new Abstract: LLM serving is increasingly accelerated by position-independent caching (PIC).
The paper introduces LatentPort, a method that allows a language model to transfer its live memory to another model without requiring the receiver to reread the context. Experiments on a Qwen3.5 4B-to-9B sibling pair show that adding a Gated DeltaNet (GDN) persistent-state package reduces negative log‑likelihood by 0.747 nats/token and improves performance across 64 PG19 documents. The study also demonstrates that direct recurrent and convolution reuse outperforms learned GDN maps, and a 434,176‑parameter correction further narrows the performance gap to the native 9B model.
RaReCache is a framework that allows a large target language model to decode accurately from a KV cache prefilled by a much smaller source model by selectively recomputing only critical token positions. It identifies these positions using a rank disagreement metric that scores tokens based on the energy of their mapped KV in output directions weakly supported by calibration data. Across two model families and five benchmarks, RaReCache achieves up to a 3.04× prefill speedup and significantly reduces prefill latency while maintaining 95–99% of target accuracy with a 30–40% recompute budget.
CacheBridge is a method for efficiently transferring key‑value (KV) caches between large language models (LLMs) in a multi‑model system. It replaces the full‑head mapping approach by matching each target KV head to a single source head, weighting reconstruction errors by causal attention sensitivity, and building weighted sufficient statistics with a fused GPU kernel. The technique achieves comparable or better accuracy to full‑head mapping while reducing mapper storage by up to eight‑fold, accelerating application by up to three‑times, and cutting construction time dramatically.
arXiv:2609.33051v2 Announce Type: replace Abstract: Hybrid-attention models replace most softmax attention layers with linear attention, reducing KV-cache growth and enabling larger decode batches wh...
arXiv:2608.30386v1 Announce Type: cross Abstract: Hybrid linear-attention architectures have recently scaled to large open-weight models, offering quality competitive with full attention while substa...
arXiv:2608.30310v1 Announce Type: cross Abstract: Hybrid large language models interleave full-attention layers with linear-attention layers to reduce the cost of long-context inference. This structu...
CacheReforge is a method for recovering stale key‑value (KV) caches in large language models when lightweight adapters evolve. It represents stale caches as layer‑wise mixed‑version objects and uses adapter anchors, sensitivity calibration, drift accumulation, and restart boundaries to decide between direct reuse, bounded recomputation, or full suffix recovery. Experiments on Qwen2.5 models with continual LoRA updates show a 92.4% reduction in mean KL divergence while only recomputing 5.44% of layers and cutting cache‑maintenance time by 93.2% compared to full prefill.
arXiv:2608. 03893v1 Announce Type: new Abstract: Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch.