analiza aiqbench
SelKV introduces a training-free framework that optimizes KV cache compression by adaptively merging tokens based on value-vector similarity while correcting for resulting attention imbalances. This approach significantly reduces memory footprint to 25% of the original size without degrading generation quality, offering substantial speedups for long-context inference tasks. The method represents a notable advancement in making large language models more efficient and scalable through targeted cache management techniques.
SelKV: Selective KV Cache Merging with Per-Token Merge-or-Drop and Attention Compensation
arXiv:2607.16213v1 Announce Type: new Abstract: Large Language Models (LLMs) generate text autoregressively, relying on a key-value (KV) cache whose memory footprint grows linearly with context length, creating a major bottleneck. Recent compression methods mitigate this cost via token merging; however, these approaches often rely on indiscriminate aggregation, which degrades representations and introduces attention sag, a mismatch where merged tokens receive the same softmax mass as individual tokens despite encoding multiple inputs. We propose a training-free, dual-component framework for KV cache compression that addresses these limitations. First, a soft cosine gate adaptively modulates merging decisions based on value-vector similarity, suppressing or discarding dissimilar tokens to preserve semantic fidelity. Second, we introduce an attention-ratio compensation mechanism that applies a decoding-time logit bias derived from prefill attention statistics, correcting the softmax imbalance induced by merging. Evaluated on LongBench (16 English datasets) while retaining only 25% of the KV cache, our framework achieves strong compressed performance against representative one-shot baselines. It is especially robust on the evaluated grouped-query attention (GQA) models, maintaining nearlossless generation quality. Furthermore, the method outperforms the full-cache baseline on complex multi-document QA tasks and delivers a 3.3x decoding speedup at 100k tokens.
Abstract:Large Language Models (LLMs) generate text autoregressively, relying on a key-value (KV) cache whose memory footprint grows linearly with context length, creating a major bottleneck. Recent compression methods mitigate this cost via token merging; however, these approaches often rely on indiscriminate aggregation, which degrades representations and introduces attention sag, a mismatch where merged tokens receive the same softmax mass as individual tokens despite encoding multiple inputs. We propose a training-free, dual-component framework for KV cache compression that addresses these limitations. First, a soft cosine gate adaptively modulates merging decisions based on value-vector similarity, suppressing or discarding dissimilar tokens to preserve semantic fidelity. Second, we introduce an attention-ratio compensation mechanism that applies a decoding-time logit bias derived from prefill attention statistics, correcting the softmax imbalance induced by merging. Evaluated on LongBench (16 English datasets) while retaining only 25% of the KV cache, our framework achieves strong compressed performance against representative one-shot baselines. It is especially robust on the evaluated grouped-query attention (GQA) models, maintaining nearlossless generation quality. Furthermore, the method outperforms the full-cache baseline on complex multi-document QA tasks and delivers a 3.3x decoding speedup at 100k tokens.
Artificial Intelligence (cs.AI)
arXiv:2607.16213 [cs.AI]
(or arXiv:2607.16213v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2607.16213 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 279 cuvinte
Subiecte conexe
[AINews] not much happened today
Latent Space · 3 oct. 2026
[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
Latent Space · 13 oct. 2026
[AINews] The Future of Latent Space
Latent Space · 25 sept. 2026
Benchmarking Kimi-K3 Across vLLM, SGLang, and ATOM on MI350X
AMD · 22 sept. 2026