惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

B
Blog
Hugging Face - Blog
Hugging Face - Blog
月光博客
月光博客
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
人人都是产品经理
人人都是产品经理
博客园 - Franky
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
P
Proofpoint News Feed
F
Fortinet All Blogs
H
Help Net Security
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
V
Visual Studio Blog
Jina AI
Jina AI
J
Java Code Geeks
Blog — PlanetScale
Blog — PlanetScale
S
SegmentFault 最新的问题
D
DataBreaches.Net
T
The Blog of Author Tim Ferriss
美团技术团队
博客园 - 司徒正美
宝玉的分享
宝玉的分享
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
Apple Machine Learning Research
Apple Machine Learning Research

cs.NE updates on arXiv.org

MPCS: Neuroplastic Continual Learning via Multi-Component Plasticity and Topology-Aware EWC Combining Trained Models in Reinforcement Learning Training Non-Differentiable Networks via Optimal Transport ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization Probe-Geometry Alignment: Erasing the Cross-Sequence Memorization Signature Below Chance Benchmarking local Hebbian learning rules for memory storage and prototype extraction Robust volatility updates for Hierarchical Gaussian Filtering Spiking Sequence Machines and Transformers Affinity Is Not Enough: Recovering the Free Energy Principle in Mixture-of-Experts Scalable Learning in Structured Recurrent Spiking Neural Networks without Backpropagation Geometric and dynamical analysis of attractor boundaries and storage limits in kernel Hopfield networks Attractor FCM Physical Foundation Models: Fixed hardware implementations of large-scale neural networks When Does Structure Matter in Continual Learning? Dimensionality Controls When Modularity Shapes Representational Geometry Learning to Forget: Continual Learning with Adaptive Weight Decay Causal Learning with Neural Assemblies NORACL: Neurogenesis for Oracle-free Resource-Adaptive Continual Learning Text-Utilization for Encoder-dominated Speech Recognition Models EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures EvoTSC: Evolving Feature Learning Models for Time Series Classification via Genetic Programming Analysis and Explainability of LLMs Via Evolutionary Methods Deployment-Aligned Low-Precision Neural Architecture Search for Spaceborne Edge AI SeaEvo: Advancing Algorithm Discovery with Strategy Space Evolution Primitive Recursion without Composition: Dynamical Characterizations, from Neural Networks to Polynomial ODEs MAEO: Multiobjective Animorphic Ensemble Optimization for Scalable Large-scale Engineering Applications Necessary and sufficient conditions for universality of Kolmogorov-Arnold networks Learn&Drop: Fast Learning of CNNs based on Layer Dropping Architecture-Induced Recoverability Bias in Differentiable Symbolic Regression Collocation-based Robust Physics Informed Neural Networks for time-dependent simulations of pollution propagation under thermal inversion conditions on Spitsbergen Structure-Guided Diffusion Model for EEG-Based Visual Cognition Reconstruction
KV Cache Recycling to Expand Usable Context Capacity in L...
Prashant Pandey · 2025-12-05 · via cs.NE updates on arXiv.org

Whether attention key value (KV) states computed for one prompt for a small LLM can be reused to accelerate inference on a new similar prompt, giving an increase to the space to its context memory using an approach called token recycling. Using a standard Hugging Face setup with DialoGPT-medium (a 345M parameter GPT-2 style decoder trained on 147M Reddit exchanges, 2005 to 2017) as the testbed, we build a cache of past activations and get entries by sentence embeddings, then reuse cached past key values when the cached prompt is an exact prefix of the new input. We compare recycled vs. baseline runs on latency and output fidelity, and log reuse depth in tokens. Reproducibility requires no model modifications, cached KVs are serialized to the CPU, reloaded, and supplied to the generate function to continue decoding from the cached prefix. In tests, we observe consistent speedups when prefix overlap exists, with no material degradation in output semantics, and when overlap is absent, behavior matches baseline.