Close Enough is Nowhere Near: Fidelity-Consumability Decoupling in KV Caches
We show that geometric similarity to that cache does not predict whether the reader can use it, and the failure is not gradual, so a more faithful substitute is not the remedy it appears to be.
ProofPaper ↗
Key points
- Multi-agent LLM systems communicate in text, which is lossy and wasteful: the sender throws away the reading it already did.
- A growing literature instead transmits the KV cache, the natural carrier of that reading.
- We study the most appealing form of the idea: both models stay frozen and the sender's cache is substituted for the reader's with nothing trained.
- Experimentally, we test eight diverse substitution methods against two measured anchors: an uninformative cache baseline and the reader's own accuracy from having read the passage.
Sources (1)
- [1]Close Enough is Nowhere Near: Fidelity-Consumability Decoupling in KV CachesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 02:56 PM
We show that geometric similarity to that cache does not predict whether the reader can use it, and the failure is not gradual, so a more faithful substitute is not the remedy it appears to be.
Multi-agent LLM systems communicate in text, which is lossy and wasteful: the sender throws away the reading it already did.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026SUPRA-TEX: 3D Generation with Super-Aligned Texture
- Oct 15, 2026When the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel Prefill
- Oct 14, 2026mistralai/Timestral-209M
- Oct 14, 2026From Failures to Learnable Worlds: Policy-Conditioned Environment Generation for Navigation Agents
- Oct 14, 2026VLX-Go: Robust Embodied Tracking through Multi-Stage Learning and On-Policy Refinement
- Oct 14, 2026StreamErase: Towards Ultra-Fast Streaming Video Erasing