analiza aiqbench
This research investigates how sigmoid attention mechanisms can resolve the soft-to-hard mismatch in learned KV-cache eviction, enabling efficient memory management without significant performance loss. The study demonstrates that specific attention normalization choices are critical for transferring training-time gating strategies to effective inference-time deletion protocols. These findings offer a concrete architectural improvement for optimizing large language model inference efficiency.
Sigmoid Attention as a Better Substrate for Learned KV Cache Eviction
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 inference saves memory only when KV entries are physically removed. We ask whether the attention substrate affects this soft-to-hard transition. Using GPT-2-scale Transformers trained on OpenWebText, we run a controlled $2\times2\times2$ comparison over attention type, learned gating, and positional encoding. Although sigmoid attention is worse as a dense language model, learned hard eviction changes the useful operating points: sigmoid-gated models delete KV entries with negligible PPL change relative to their own no-eviction references. Under a matched live-cache protocol on the same dense backbones, learned sigmoid gates obtain lower PPL than our H$_2$O and KeyDiff implementations, whereas softmax gates do not uniformly beat these post-hoc methods. The results suggest that attention normalization can substantially affect whether a training-time soft gate transfers cleanly to hard KV deletion.
Abstract: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 physically removed. We ask whether the attention substrate affects this soft-to-hard transition. Using GPT-2-scale Transformers trained on OpenWebText, we run a controlled $2\times2\times2$ comparison over attention type, learned gating, and positional encoding. Although sigmoid attention is worse as a dense language model, learned hard eviction changes the useful operating points: sigmoid-gated models delete KV entries with negligible PPL change relative to their own no-eviction references. Under a matched live-cache protocol on the same dense backbones, learned sigmoid gates obtain lower PPL than our H$_2$O and KeyDiff implementations, whereas softmax gates do not uniformly beat these post-hoc methods. The results suggest that attention normalization can substantially affect whether a training-time soft gate transfers cleanly to hard KV deletion.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2608.23296 [cs.LG]
(or arXiv:2608.23296v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2608.23296 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
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 · 239 cuvinte
Subiecte conexe
Productionizing TurboQuant on AMD GPUs for KV-Cache-Bound LLM Inference
AMD · 11 iun. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
Error Certificates for KV-Cache Eviction via Randomized Design
arXiv · 25 iul. 2026
ARM: Attention with Routed-Memory for Learnable Sparse Control
arXiv · 22 sept. 2026