#kv cache

try —

6 papers match

cs.LG2026

Stage-Replay Divergence Follows the KV Cache: Fixed-Prefix Precision Controls and Bidirectional Cache Transplantation

Alexander Boesgaard Lorup

The paper investigates how the key/value cache in transformer language models affects token generation, showing that cache contents alone can reproduce divergent outputs and that n…

#transformer caching#kv cache#precision effects#stage replay
cs.DC2026

DualDecoder: Accelerate Long Context LLM Inference by Predictive Prefetch

Zuning Liang, Zhiyi Yao, Qi Chen +6

DualDecoder is a serving system that predicts and prefetches the key‑value cache entries needed for the next token in long‑context LLM inference, reducing GPU memory overhead and b…

#long-context inference#kv cache#sparse retrieval#predictive prefetch
cs.PF2026

A Photonic-CXL Memory Appliance for Scalable KV Cache Management in LLM Inference

Jing Ding, Yash Nishant, Chandrish Ambati +2

The paper proposes a photonic‑CXL memory appliance that uses a passive fiber shuffle to provide a 32 TB shared memory pool for large‑language‑model KV cache management, reducing la…

#photonic interconnects#cxl memory#kv cache#llm inference
cs.DC2026

InferScale: GPU-Native KV Injection for Personalized LLM Serving

Peter Li, Prashant Pandey

The paper introduces InferScale, a GPU-native system that precomputes and caches key‑value (KV) representations of personalized memory facts for large language models, allowing dir…

#large language models#kv cache#personalized memory#inference optimization
cs.CL2026

PReM: Learning What to Preserve and When to Refresh for Context Compression

Bohan Yu, Lei Shen, Chenxi Zhou +5

The paper proposes PReM, a framework that lets language models dynamically decide which parts of a long context to keep and when to refresh stored information, improving efficiency…

#context compression#long-context inference#memory selection#kv cache
cs.DC2026

[AAFLOW+] Stateful Operator Abstraction with Zero-Copy Distributed KV Cache Orchestration for Multi-Agent Workflows

Arup Kumar Sarker, Alexander James Halpern, Mills Staylor +5

The paper presents AAFLOW+, a framework that treats key‑value (KV) caches as distributed objects, enabling zero‑copy sharing of model state across multi‑agent LLM workflows to cut…

#large language models#multi-agent systems#kv cache#zero-copy