Abstract
We study the rate-distortion limits of online KV cache compression in autoregressive language models, formulating it as sequential Wyner-Ziv source coding on the filtration induced by the model, with the next-step query as decoder side information. Empirically, across four models spanning two families and 0.5-3B parameters, we find that the next-token distribution's sensitivity to context truncation decays \emph{polynomially} rather than \emph{geometrically}: a power law improves on an exponential fit by an order of magnitude in extrapolation, the fitted exponent is recovered independently from a sink-plus-recent KL measurement, and the decay is verified to be free of positional-encoding artifacts by a position-preserving ablation. Under a corresponding \emph{polynomial truncation-sensitivity} assumption, our main result characterizes the per-token memory requirement of \emph{suffix-only} cache policies: a sliding-window scheme attains distortion ε with window w=O(ε−1/α), and -- under an additional two-sided Bayes-risk condition -- a converse shows w=Ω(ε−1/α) is necessary within this policy class, so the scaling is Θ(ε−1/α) for suffix-only policies. Whether recurrent or propagating cache summaries can beat this scaling is left open. An explicit block-Markov scheme achieves the upper bound; its rate-of-convergence exponent matches the converse under additional forward-decay and regularity hypotheses (not implied by truncation sensitivity alone), and differs by a factor of two otherwise. Empirically, the polynomial law predicts the degradation curves of concrete cache policies: recency-based eviction (sliding, sink-plus-recent) suppresses distortion by roughly two orders of magnitude over random retention at equal budget, with a power-law decay in the budget.
Explore similar work
May 8, 2026cs.LG
Large language models (LLMs) have shown strong performance across diverse tasks, but their inference with long input contexts is bottlenecked by memory size and bandwidth. The Key-Value (KV) cache size grows linearly with sequence length and needs to be re-read from off-chip high-bandwidth memory (HBM) to on-chip memory at every decoding step, resulting in memory-bound inference. Existing methods reduce the cache by either eviction or quantization, but typically treat the two in isolation. In this paper, we cast KV cache compression as a rate-distortion problem, under which eviction and quantization are two end-points of the same bit allocation scheme. This exposes the need to optimize them jointly, motivating our method, RDKV (Rate-Distortion KV cache compression). RDKV derives the weight of each token or channel from the distortion that compression induces on the attention computation. Based on these weights, it assigns each token or channel a bit-width ranging from full precision down to zero bits guided by reverse water-filling, applied once after the prefilling stage. Experiments on LongBench, RULER, and InfiniteBench show that RDKV outperforms the best evaluated baseline by 9.1% on average. On LongBench it recovers 97.81% of full-cache accuracy with only 2.48% cache retention. Compared with full-cache FlashAttention-2 decoding, it achieves 4.5x decode speedup and 1.9x peak memory reduction with 128K context length, while maintaining comparable performance.
Junkai Zhang, Hang Guo, Luca Benini +1
Jun 13, 2026cs.LG
KV cache compression is essential for reducing the memory cost of long-context large language model inference. Existing approaches, however, typically apply a single compression policy and a uniform cache budget across all transformer layers. This uniform design ignores the fact that different layers can play different roles during prefill and decoding, and may therefore require different eviction strategies and cache capacities. We present PolyKV, a layer-wise KV cache optimization framework that considers design space with method selection and budget allocation. PolyKV routes each layer to a suitable KV compression policy based on layer-level signals, while assigning non-uniform budgets under a fixed total budget. This formulation enables heterogeneous compositions of existing KV cache methods. Experiments on LLaMA-3.1-8B and Qwen3-8B show that, under the same 512-token average KV budget, PolyKV recovers 54.5% and 25.7% of the LongBench performance gap between the strongest single-policy baseline and FullKV, respectively. Across 128-1024 budget sweep, PolyKV consistently improves over the strongest baseline by 1.7%-6.4%, corresponding to 40.0%-54.5% recovery of the FullKV gap.
Chao Fei, Panos Kalnis
Jul 1, 2026cs.LG
Transformer inference on long sequences is expensive because softmax attention repeatedly reads from a large KV cache. The prevalent approach to this bottleneck is KV cache compression, which replaces the full cache with a compact summary. Despite its practical importance, the design of such summaries is largely driven by empirical experimentation. On the theoretical side, existing results show that KV cache compression can be impossible in the worst case, but offer little systematic guidance for designing algorithms in regimes where accurate compression is possible. We bridge this gap by characterizing the minimax risk of KV cache compression in terms of the intrinsic compressibility of a cache, revealing when and how accurate compression is possible. These results yield novel design principles for KV cache compression under causal masking that map efficiently to prefill and autoregressive decoding while achieving minimax-optimal risk. We instantiate these principles in a practical algorithm and report promising performance on LongBench in targeted experiments. Overall, our results provide a principled avenue for practical KV cache compression with theoretical guarantees.
Lukas Haverbeck, Carmen Amo Alonso, Andres Felipe Posada-Moreno +2