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.