Polynomial Context-Truncation Sensitivity in Autoregressive Language Models: Sequential Wyner-Ziv Bounds for KV Cache Compression

Context truncation sensitivity in autoregressive LMs (0.5B–3B parameters across two model families) follows a power law rather than exponential decay, and under this polynomial sensitivity assumption, sliding-window KV cache schemes achieve the rate-distortion optimal per-token memory bound derived via sequential Wyner-Ziv coding theory.