CacheBridge (arXiv 2609.00891, Wed 2 Sep digest) avoids shared-prefix replay when routing requests between LLMs: it co-designs architecture-indexed mapper support, attention-aligned calibration, and bounded mapper construction while retaining a closed-form affine-mapper interface that translates source-model KV caches into target-model caches online. It improves on Full-Head Mapping, whose per-head-from-every-head design is sensitive to architectural differences and whose mapper cost grows with layer support. The same window carries an independent companion: 'A Universal Context-Reuse Layer for Cross-Model KV Sharing' (2608.30963, Tue 1 Sep digest), which studies cross-model KV translation across differences in scale, architecture, attention configuration, tokenizer, and model family.
CacheBridge transfers KV caches across different model architectures
CacheBridge (arXiv 2609.00891, Wed 2 Sep digest) avoids shared-prefix replay when routing requests between LLMs: it co-designs architecture-indexed mapper support, attention-aligned calibration, and bounded mapper construction while retaining a closed-form affine-mapper interface that translates source-model KV caches into target-model caches online. It improves on Full-Head Mapping, whose per-head-from-every-head design is sensitive to architectural differences and whose mapper cost grows with layer support. The same window carries an independent companion: 'A Universal Context-Reuse Layer for Cross-Model KV Sharing' (2608.30963, Tue 1 Sep digest), which studies cross-model KV translation across differences in scale, architecture, attention configuration, tokenizer, and model family.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
Multi-model pipelines (router -> small model, draft -> verifier, generalist -> specialist) currently pay full prefill on every handoff, which dominates cost exactly where routing is supposed to save money. Two independent groups publishing cross-model KV transfer in the same window suggests the assumption 'KV caches are model-specific' is starting to bend; teams running model-routing stacks should track this the way they tracked prompt caching in 2025.
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Arxiv | 2609.00891, announced in the Wed 2 Sep 2026 digest |
|---|---|
| Mechanism | closed-form affine mapper from source to target KV caches, online-deployable; architecture-indexed mapper support + attention-aligned calibration + bounded mapper construction |
| Improves On | Full-Head Mapping (maps each target head from every source head in selected layers): transfer quality sensitive to architectural differences; mapper storage/application cost grows with layer support |
| Companion | arXiv 2608.30963 'A Universal Context-Reuse Layer for Cross-Model KV Sharing' (Tue 1 Sep digest): cross-model KV translation across scale/architecture/attention/tokenizer/family differences; evaluated within-family (Qwen2.5-7B -> Qwen2.5-1.5B) and cross-family |
| Kb Thread | continues the KV-reuse line: ReCache (2608.19662, 8/21, composition-invariant reuse for tool-augmented agents), TwinKV (2608.27128, 8/29, eviction repair) |