arXiv:2610.14754v1 Announce Type: new Abstract: Hybrid language models interleave Gated DeltaNet (GDN) layers, whose recurrent memory has a fixed size, with full-attention layers, whose key-value (KV...
ValueDiff introduces a value‑geometric KV cache eviction strategy for large language models that suppress attention sinks. It ranks tokens by the L2 deviation of their value vectors from the cache mean, a score that aligns with minimal‑disturbance eviction under a max‑entropy assumption. Across several benchmarks—RULER, LongBench, and MATH‑500—ValueDiff consistently retains a higher proportion of useful tokens than prior methods, especially under tight cache budgets.
The paper introduces Random Attention, a method that evicts KV cache entries uniformly at random within each attention head while preserving the prompt. Experiments on four models and six reasoning tasks show that this simple strategy matches the performance of the best existing eviction methods and achieves 32‑43% higher throughput in vLLM deployments. The authors explain that the prompt is the most fragile cache component and that reasoning traces are redundantly stored across text and attention heads, making a selection score unnecessary.
PAGE is a partition‑aware gated KV‑cache eviction method that reframes eviction as a per‑input admission decision. It uses a single label‑free scalar— the early‑to‑late drop in pairwise top‑k head agreement—to classify inputs into a capacity‑bound class (where eviction is catastrophic) and a dilution‑prone class (where eviction is safe or beneficial). By thresholding this drop, PAGE applies a base evictor only when necessary, reducing the harm rate in the capacity‑bound regime from 0.75 to 0.026 and achieving a 29× improvement across four models and benchmarks without retraining the evictor.
The paper introduces Random Attention, a method that evicts KV cache entries uniformly at random within each attention head while preserving the prompt. It demonstrates that this simple strategy matches or surpasses more complex eviction schemes across four models and six reasoning tasks, achieving 32‑43% higher throughput in vLLM deployments. Experiments reveal that the prompt is the most fragile cache component and that redundancy in the reasoning trace across text and attention heads protects against random eviction, eliminating the need for a selection score.
arXiv:2608.23296v1 Announce Type: cross Abstract: Learned KV-cache eviction often faces a soft-to-hard mismatch: during training, differentiable gates typically attenuate token contributions, whereas...
AbridgeKV compresses long chain‑of‑thought outputs by encoding completed text blocks into the KV cache of specially trained pause tokens, freeing the original cache space. The method uses pause tokens positioned within the block, allocates cache budget per layer based on entry value, and compresses large blocks. Experiments on 446 competition problems show that with only a quarter of the original KV cache, AbridgeKV matches the accuracy and output length of a full‑KV model, while enabling faster inference and a 3.5× smaller cache per sequence.
arXiv:2608. 11231v1 Announce Type: new Abstract: LLM serving is increasingly accelerated by position-independent caching (PIC).
arXiv:2610.07643v1 Announce Type: cross Abstract: Most KV-cache eviction methods ask, in effect, which memory appeared important while reading the prompt? We instead ask, which memory will matter whi...
The paper introduces Draft‑Guided Eviction (DGE), a training‑free KV‑cache compression technique that postpones token eviction until after drafting the first two answer tokens, thereby preserving the cache entries needed for the answer trajectory. Unlike prior methods that focus on optimizing what to keep, DGE changes the timing of eviction, allowing existing compression schemes such as SnapKV, H2O, and PyramidKV to be applied without altering their eviction scores. Experiments show DGE consistently outperforms previous approaches across multiple instruct‑tuned models, achieving near‑full‑cache performance on LongBench and demonstrating that the improvement stems from trajectory anchoring rather than changes in cache selection.
Learned KV-cache eviction often faces a soft-to-hard mismatch: during training, differentiable gates typically attenuate token contributions, whereas inference saves memory only when KV entries are ph...
The paper investigates how block‑diffusion language models can use a constant‑size cache to enable efficient parallel decoding. By employing sequence mixers that summarize completed blocks into a reusable state and a block‑causal training objective, the authors pretrain three 3B block‑diffusion denoisers (attention, Mamba, and hybrid) on 300 B tokens. The resulting state‑space cache remains O(1) in memory and latency regardless of context length, yielding significant speed‑up and memory savings compared to traditional attention‑based caches, especially at very long sequences.