Task-Level Concealment for KV-Cache Transmission

Develop task-level error-concealment mechanisms for perturbed key-value cache tensors transmitted between large-model endpoints, so that cache corruption can be mitigated without relying solely on automatic repeat request or reduced coding rates.

Background

Token identifier transmission can exploit a receiver’s generative prior to conceal index errors, whereas key-value cache transmission exposes perturbed tensor entries directly to the attention computation. The paper identifies the absence of task-level concealment for corrupted KV-cache tensors as an unresolved issue in selecting between token forwarding and cache transfer as communication media for distributed large-model inference.

References

Task-level concealment remains an open problem.

— From Semantic to Token Communication: The Next Paradigm for Large-Model-Driven 6G Intelligent Connectivity  (2609.10714 - Ma et al., 9 Sep 2026) in Table 10, Section 5.6 (Token Communication for LM Services)

The open question is therefore not how to make a produced cache closer, but how to confine its residual error to directions the reader cannot resolve, and nothing in our sweep suggests the sender's side has the information required.

— Close Enough is Nowhere Near: Fidelity-Consumability Decoupling in KV Caches  (2610.22016 - Deng et al., 15 Oct 2026) in Section 4, “Why Substitution Cannot Reach the Tolerable Set,” subsection “The Softmax Consumption Point”