AlignPred: Error-Controlled KV Cache Compression via Alignment and Predictive Coding
We present AlignPred, a codec that combines local alignment with predictive coding under a configurable KV reconstruction-error budget.
ProofPaper ↗
Key points
- Long-running agents retain KV caches across turns and sessions for hours or days.
- Compressing these persistent caches requires reducing storage while controlling reconstruction error.
- AlignPred achieves near-native task scores at substantially compressed operating points on LongBench-Pro at 128K context.
- AlignPred's SGLang HiCache integration reduces persistent target-KV storage by 42-72% relative to native HiCache, with comparable latency.
Sources (1)
- [1]AlignPred: Error-Controlled KV Cache Compression via Alignment and Predictive CodingarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 08:44 AM
We present AlignPred, a codec that combines local alignment with predictive coding under a configurable KV reconstruction-error budget.
Long-running agents retain KV caches across turns and sessions for hours or days.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026Fish Audio Drama 3 Technical Report
- Oct 12, 2026Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents
- Oct 12, 2026Hashrope: A Persistent Data Structure for Mechanisms of LLM Context Management
- Oct 12, 2026DuoDiscover: Scaling Autonomous Discovery Beyond the Evaluation Bottleneck
- Oct 11, 2026Agents That Own Their Inference — Du'an Lightfoot & Khaja Omer, Akamai Technologies
- Oct 11, 2026UPDATE: Qwen 3.8 27B 140 tok/s on single RTX 3090 Megakernel: KL divergence 0.0009 vs llama.cpp